1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86TargetMachine.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/GlobalAlias.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/ADT/BitVector.h"
27 #include "llvm/ADT/VectorExtras.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineModuleInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/PseudoSourceValue.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Target/TargetOptions.h"
37 #include "llvm/ADT/SmallSet.h"
38 #include "llvm/ADT/StringExtras.h"
39 #include "llvm/Support/CommandLine.h"
43 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
45 // Forward declarations.
46 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
49 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
50 : TargetLowering(TM) {
51 Subtarget = &TM.getSubtarget<X86Subtarget>();
52 X86ScalarSSEf64 = Subtarget->hasSSE2();
53 X86ScalarSSEf32 = Subtarget->hasSSE1();
54 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
56 RegInfo = TM.getRegisterInfo();
59 // Set up the TargetLowering object.
61 // X86 is weird, it always uses i8 for shift amounts and setcc results.
62 setShiftAmountType(MVT::i8);
63 setBooleanContents(ZeroOrOneBooleanContent);
64 setSchedulingPreference(SchedulingForRegPressure);
65 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
66 setStackPointerRegisterToSaveRestore(X86StackPtr);
68 if (Subtarget->isTargetDarwin()) {
69 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
70 setUseUnderscoreSetJmp(false);
71 setUseUnderscoreLongJmp(false);
72 } else if (Subtarget->isTargetMingw()) {
73 // MS runtime is weird: it exports _setjmp, but longjmp!
74 setUseUnderscoreSetJmp(true);
75 setUseUnderscoreLongJmp(false);
77 setUseUnderscoreSetJmp(true);
78 setUseUnderscoreLongJmp(true);
81 // Set up the register classes.
82 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
83 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
84 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
85 if (Subtarget->is64Bit())
86 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
88 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
90 // We don't accept any truncstore of integer registers.
91 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
92 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
93 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
94 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
95 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
96 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
98 // SETOEQ and SETUNE require checking two conditions.
99 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
100 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
101 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
102 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
103 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
104 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
106 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
108 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
109 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
110 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
112 if (Subtarget->is64Bit()) {
113 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
114 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
115 } else if (!UseSoftFloat) {
116 if (X86ScalarSSEf64) {
117 // We have an impenetrably clever algorithm for ui64->double only.
118 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
120 // We have an algorithm for SSE2, and we turn this into a 64-bit
121 // FILD for other targets.
122 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
125 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
127 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
128 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
131 // SSE has no i16 to fp conversion, only i32
132 if (X86ScalarSSEf32) {
133 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
134 // f32 and f64 cases are Legal, f80 case is not
135 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
137 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
138 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
141 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
142 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
145 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
146 // are Legal, f80 is custom lowered.
147 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
148 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
150 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
152 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
153 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
155 if (X86ScalarSSEf32) {
156 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
157 // f32 and f64 cases are Legal, f80 case is not
158 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
160 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
161 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
164 // Handle FP_TO_UINT by promoting the destination to a larger signed
166 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
167 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
168 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
170 if (Subtarget->is64Bit()) {
171 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
172 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
173 } else if (!UseSoftFloat) {
174 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
175 // Expand FP_TO_UINT into a select.
176 // FIXME: We would like to use a Custom expander here eventually to do
177 // the optimal thing for SSE vs. the default expansion in the legalizer.
178 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
180 // With SSE3 we can use fisttpll to convert to a signed i64; without
181 // SSE, we're stuck with a fistpll.
182 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
185 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
186 if (!X86ScalarSSEf64) {
187 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
188 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
191 // Scalar integer divide and remainder are lowered to use operations that
192 // produce two results, to match the available instructions. This exposes
193 // the two-result form to trivial CSE, which is able to combine x/y and x%y
194 // into a single instruction.
196 // Scalar integer multiply-high is also lowered to use two-result
197 // operations, to match the available instructions. However, plain multiply
198 // (low) operations are left as Legal, as there are single-result
199 // instructions for this in x86. Using the two-result multiply instructions
200 // when both high and low results are needed must be arranged by dagcombine.
201 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
202 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
203 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
204 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
205 setOperationAction(ISD::SREM , MVT::i8 , Expand);
206 setOperationAction(ISD::UREM , MVT::i8 , Expand);
207 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
208 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
209 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
210 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
211 setOperationAction(ISD::SREM , MVT::i16 , Expand);
212 setOperationAction(ISD::UREM , MVT::i16 , Expand);
213 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
214 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
215 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
216 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
217 setOperationAction(ISD::SREM , MVT::i32 , Expand);
218 setOperationAction(ISD::UREM , MVT::i32 , Expand);
219 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
220 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
221 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
222 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
223 setOperationAction(ISD::SREM , MVT::i64 , Expand);
224 setOperationAction(ISD::UREM , MVT::i64 , Expand);
226 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
227 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
228 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
229 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
230 if (Subtarget->is64Bit())
231 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
232 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
233 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
234 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
235 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
236 setOperationAction(ISD::FREM , MVT::f32 , Expand);
237 setOperationAction(ISD::FREM , MVT::f64 , Expand);
238 setOperationAction(ISD::FREM , MVT::f80 , Expand);
239 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
241 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
242 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
243 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
244 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
245 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
246 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
247 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
248 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
249 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
250 if (Subtarget->is64Bit()) {
251 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
252 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
253 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
256 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
257 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
259 // These should be promoted to a larger select which is supported.
260 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
261 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
262 // X86 wants to expand cmov itself.
263 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
264 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
265 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
266 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
267 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
268 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
269 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
270 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
271 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
272 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
273 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
274 if (Subtarget->is64Bit()) {
275 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
276 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
278 // X86 ret instruction may pop stack.
279 setOperationAction(ISD::RET , MVT::Other, Custom);
280 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
283 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
284 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
285 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
286 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
287 if (Subtarget->is64Bit())
288 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
289 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
290 if (Subtarget->is64Bit()) {
291 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
292 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
293 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
294 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
296 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
297 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
298 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
299 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
300 if (Subtarget->is64Bit()) {
301 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
302 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
303 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
306 if (Subtarget->hasSSE1())
307 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
309 if (!Subtarget->hasSSE2())
310 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
312 // Expand certain atomics
313 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
314 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
315 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
316 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
318 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
319 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
320 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
321 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
323 if (!Subtarget->is64Bit()) {
324 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
325 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
326 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
327 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
328 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
329 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
330 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
333 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
334 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
335 // FIXME - use subtarget debug flags
336 if (!Subtarget->isTargetDarwin() &&
337 !Subtarget->isTargetELF() &&
338 !Subtarget->isTargetCygMing()) {
339 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
340 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
343 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
344 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
345 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
346 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
347 if (Subtarget->is64Bit()) {
348 setExceptionPointerRegister(X86::RAX);
349 setExceptionSelectorRegister(X86::RDX);
351 setExceptionPointerRegister(X86::EAX);
352 setExceptionSelectorRegister(X86::EDX);
354 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
355 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
357 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
359 setOperationAction(ISD::TRAP, MVT::Other, Legal);
361 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
362 setOperationAction(ISD::VASTART , MVT::Other, Custom);
363 setOperationAction(ISD::VAEND , MVT::Other, Expand);
364 if (Subtarget->is64Bit()) {
365 setOperationAction(ISD::VAARG , MVT::Other, Custom);
366 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
368 setOperationAction(ISD::VAARG , MVT::Other, Expand);
369 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
372 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
373 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
374 if (Subtarget->is64Bit())
375 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
376 if (Subtarget->isTargetCygMing())
377 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
379 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
381 if (!UseSoftFloat && X86ScalarSSEf64) {
382 // f32 and f64 use SSE.
383 // Set up the FP register classes.
384 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
385 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
387 // Use ANDPD to simulate FABS.
388 setOperationAction(ISD::FABS , MVT::f64, Custom);
389 setOperationAction(ISD::FABS , MVT::f32, Custom);
391 // Use XORP to simulate FNEG.
392 setOperationAction(ISD::FNEG , MVT::f64, Custom);
393 setOperationAction(ISD::FNEG , MVT::f32, Custom);
395 // Use ANDPD and ORPD to simulate FCOPYSIGN.
396 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
397 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
399 // We don't support sin/cos/fmod
400 setOperationAction(ISD::FSIN , MVT::f64, Expand);
401 setOperationAction(ISD::FCOS , MVT::f64, Expand);
402 setOperationAction(ISD::FSIN , MVT::f32, Expand);
403 setOperationAction(ISD::FCOS , MVT::f32, Expand);
405 // Expand FP immediates into loads from the stack, except for the special
407 addLegalFPImmediate(APFloat(+0.0)); // xorpd
408 addLegalFPImmediate(APFloat(+0.0f)); // xorps
409 } else if (!UseSoftFloat && X86ScalarSSEf32) {
410 // Use SSE for f32, x87 for f64.
411 // Set up the FP register classes.
412 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
413 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
415 // Use ANDPS to simulate FABS.
416 setOperationAction(ISD::FABS , MVT::f32, Custom);
418 // Use XORP to simulate FNEG.
419 setOperationAction(ISD::FNEG , MVT::f32, Custom);
421 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
423 // Use ANDPS and ORPS to simulate FCOPYSIGN.
424 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
425 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
427 // We don't support sin/cos/fmod
428 setOperationAction(ISD::FSIN , MVT::f32, Expand);
429 setOperationAction(ISD::FCOS , MVT::f32, Expand);
431 // Special cases we handle for FP constants.
432 addLegalFPImmediate(APFloat(+0.0f)); // xorps
433 addLegalFPImmediate(APFloat(+0.0)); // FLD0
434 addLegalFPImmediate(APFloat(+1.0)); // FLD1
435 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
436 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
439 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
440 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
442 } else if (!UseSoftFloat) {
443 // f32 and f64 in x87.
444 // Set up the FP register classes.
445 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
446 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
448 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
449 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
450 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
451 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
454 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
455 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
457 addLegalFPImmediate(APFloat(+0.0)); // FLD0
458 addLegalFPImmediate(APFloat(+1.0)); // FLD1
459 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
460 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
461 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
462 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
463 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
464 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
467 // Long double always uses X87.
469 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
470 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
471 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
474 APFloat TmpFlt(+0.0);
475 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
477 addLegalFPImmediate(TmpFlt); // FLD0
479 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
480 APFloat TmpFlt2(+1.0);
481 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
483 addLegalFPImmediate(TmpFlt2); // FLD1
484 TmpFlt2.changeSign();
485 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
489 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
490 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
494 // Always use a library call for pow.
495 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
496 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
497 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
499 setOperationAction(ISD::FLOG, MVT::f80, Expand);
500 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
501 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
502 setOperationAction(ISD::FEXP, MVT::f80, Expand);
503 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
505 // First set operation action for all vector types to either promote
506 // (for widening) or expand (for scalarization). Then we will selectively
507 // turn on ones that can be effectively codegen'd.
508 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
509 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
510 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
511 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
512 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
513 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
514 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
515 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
516 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
517 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
518 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
519 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
520 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
521 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
522 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
523 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
524 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
525 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
526 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
527 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
528 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
529 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
531 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
532 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
535 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
560 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
561 // with -msoft-float, disable use of MMX as well.
562 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
563 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
564 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
565 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
566 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
567 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
569 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
570 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
571 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
572 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
574 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
575 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
576 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
577 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
579 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
580 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
582 setOperationAction(ISD::AND, MVT::v8i8, Promote);
583 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
584 setOperationAction(ISD::AND, MVT::v4i16, Promote);
585 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
586 setOperationAction(ISD::AND, MVT::v2i32, Promote);
587 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
588 setOperationAction(ISD::AND, MVT::v1i64, Legal);
590 setOperationAction(ISD::OR, MVT::v8i8, Promote);
591 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
592 setOperationAction(ISD::OR, MVT::v4i16, Promote);
593 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
594 setOperationAction(ISD::OR, MVT::v2i32, Promote);
595 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
596 setOperationAction(ISD::OR, MVT::v1i64, Legal);
598 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
599 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
600 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
601 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
602 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
603 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
604 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
606 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
607 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
608 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
609 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
610 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
611 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
612 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
613 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
614 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
616 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
617 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
618 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
619 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
620 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
622 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
623 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
624 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
625 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
627 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
628 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
629 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
630 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
632 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
634 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
635 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
636 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
637 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
638 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
639 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
642 if (!UseSoftFloat && Subtarget->hasSSE1()) {
643 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
645 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
646 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
647 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
648 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
649 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
650 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
651 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
652 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
654 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
655 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
656 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
659 if (!UseSoftFloat && Subtarget->hasSSE2()) {
660 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
662 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
663 // registers cannot be used even for integer operations.
664 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
665 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
666 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
667 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
669 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
670 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
671 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
672 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
673 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
674 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
675 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
676 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
677 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
678 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
679 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
680 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
681 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
682 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
683 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
684 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
686 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
689 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
691 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
692 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
693 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
694 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
695 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
697 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
698 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
699 MVT VT = (MVT::SimpleValueType)i;
700 // Do not attempt to custom lower non-power-of-2 vectors
701 if (!isPowerOf2_32(VT.getVectorNumElements()))
703 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
704 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
705 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
708 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
709 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
711 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
712 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
713 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
715 if (Subtarget->is64Bit()) {
716 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
717 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
720 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
721 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
722 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
723 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
724 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
725 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
726 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
727 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
728 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
729 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
730 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
731 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
734 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
736 // Custom lower v2i64 and v2f64 selects.
737 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
738 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
739 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
740 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
742 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
743 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
744 if (!DisableMMX && Subtarget->hasMMX()) {
745 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
746 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
750 if (Subtarget->hasSSE41()) {
751 // FIXME: Do we need to handle scalar-to-vector here?
752 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
754 // i8 and i16 vectors are custom , because the source register and source
755 // source memory operand types are not the same width. f32 vectors are
756 // custom since the immediate controlling the insert encodes additional
758 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
759 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
760 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
761 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
763 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
764 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
765 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
766 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
768 if (Subtarget->is64Bit()) {
769 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
770 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
774 if (Subtarget->hasSSE42()) {
775 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
778 // We want to custom lower some of our intrinsics.
779 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
781 // Add/Sub/Mul with overflow operations are custom lowered.
782 setOperationAction(ISD::SADDO, MVT::i32, Custom);
783 setOperationAction(ISD::SADDO, MVT::i64, Custom);
784 setOperationAction(ISD::UADDO, MVT::i32, Custom);
785 setOperationAction(ISD::UADDO, MVT::i64, Custom);
786 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
787 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
788 setOperationAction(ISD::USUBO, MVT::i32, Custom);
789 setOperationAction(ISD::USUBO, MVT::i64, Custom);
790 setOperationAction(ISD::SMULO, MVT::i32, Custom);
791 setOperationAction(ISD::SMULO, MVT::i64, Custom);
793 if (!Subtarget->is64Bit()) {
794 // These libcalls are not available in 32-bit.
795 setLibcallName(RTLIB::SHL_I128, 0);
796 setLibcallName(RTLIB::SRL_I128, 0);
797 setLibcallName(RTLIB::SRA_I128, 0);
800 // We have target-specific dag combine patterns for the following nodes:
801 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
802 setTargetDAGCombine(ISD::BUILD_VECTOR);
803 setTargetDAGCombine(ISD::SELECT);
804 setTargetDAGCombine(ISD::SHL);
805 setTargetDAGCombine(ISD::SRA);
806 setTargetDAGCombine(ISD::SRL);
807 setTargetDAGCombine(ISD::STORE);
808 if (Subtarget->is64Bit())
809 setTargetDAGCombine(ISD::MUL);
811 computeRegisterProperties();
813 // FIXME: These should be based on subtarget info. Plus, the values should
814 // be smaller when we are in optimizing for size mode.
815 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
816 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
817 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
818 allowUnalignedMemoryAccesses = true; // x86 supports it!
819 setPrefLoopAlignment(16);
820 benefitFromCodePlacementOpt = true;
824 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
829 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
830 /// the desired ByVal argument alignment.
831 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
834 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
835 if (VTy->getBitWidth() == 128)
837 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
838 unsigned EltAlign = 0;
839 getMaxByValAlign(ATy->getElementType(), EltAlign);
840 if (EltAlign > MaxAlign)
842 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
843 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
844 unsigned EltAlign = 0;
845 getMaxByValAlign(STy->getElementType(i), EltAlign);
846 if (EltAlign > MaxAlign)
855 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
856 /// function arguments in the caller parameter area. For X86, aggregates
857 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
858 /// are at 4-byte boundaries.
859 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
860 if (Subtarget->is64Bit()) {
861 // Max of 8 and alignment of type.
862 unsigned TyAlign = TD->getABITypeAlignment(Ty);
869 if (Subtarget->hasSSE1())
870 getMaxByValAlign(Ty, Align);
874 /// getOptimalMemOpType - Returns the target specific optimal type for load
875 /// and store operations as a result of memset, memcpy, and memmove
876 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
879 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
880 bool isSrcConst, bool isSrcStr,
881 SelectionDAG &DAG) const {
882 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
883 // linux. This is because the stack realignment code can't handle certain
884 // cases like PR2962. This should be removed when PR2962 is fixed.
885 const Function *F = DAG.getMachineFunction().getFunction();
886 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
887 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
888 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
890 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
893 if (Subtarget->is64Bit() && Size >= 8)
898 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
900 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
901 SelectionDAG &DAG) const {
902 if (usesGlobalOffsetTable())
903 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
904 if (!Subtarget->isPICStyleRIPRel())
905 // This doesn't have DebugLoc associated with it, but is not really the
906 // same as a Register.
907 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
912 //===----------------------------------------------------------------------===//
913 // Return Value Calling Convention Implementation
914 //===----------------------------------------------------------------------===//
916 #include "X86GenCallingConv.inc"
918 /// LowerRET - Lower an ISD::RET node.
919 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
920 DebugLoc dl = Op.getDebugLoc();
921 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
923 SmallVector<CCValAssign, 16> RVLocs;
924 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
925 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
926 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
927 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
929 // If this is the first return lowered for this function, add the regs to the
930 // liveout set for the function.
931 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
932 for (unsigned i = 0; i != RVLocs.size(); ++i)
933 if (RVLocs[i].isRegLoc())
934 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
936 SDValue Chain = Op.getOperand(0);
938 // Handle tail call return.
939 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
940 if (Chain.getOpcode() == X86ISD::TAILCALL) {
941 SDValue TailCall = Chain;
942 SDValue TargetAddress = TailCall.getOperand(1);
943 SDValue StackAdjustment = TailCall.getOperand(2);
944 assert(((TargetAddress.getOpcode() == ISD::Register &&
945 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
946 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R11)) ||
947 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
948 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
949 "Expecting an global address, external symbol, or register");
950 assert(StackAdjustment.getOpcode() == ISD::Constant &&
951 "Expecting a const value");
953 SmallVector<SDValue,8> Operands;
954 Operands.push_back(Chain.getOperand(0));
955 Operands.push_back(TargetAddress);
956 Operands.push_back(StackAdjustment);
957 // Copy registers used by the call. Last operand is a flag so it is not
959 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
960 Operands.push_back(Chain.getOperand(i));
962 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
969 SmallVector<SDValue, 6> RetOps;
970 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
971 // Operand #1 = Bytes To Pop
972 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
974 // Copy the result values into the output registers.
975 for (unsigned i = 0; i != RVLocs.size(); ++i) {
976 CCValAssign &VA = RVLocs[i];
977 assert(VA.isRegLoc() && "Can only return in registers!");
978 SDValue ValToCopy = Op.getOperand(i*2+1);
980 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
981 // the RET instruction and handled by the FP Stackifier.
982 if (VA.getLocReg() == X86::ST0 ||
983 VA.getLocReg() == X86::ST1) {
984 // If this is a copy from an xmm register to ST(0), use an FPExtend to
985 // change the value to the FP stack register class.
986 if (isScalarFPTypeInSSEReg(VA.getValVT()))
987 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
988 RetOps.push_back(ValToCopy);
989 // Don't emit a copytoreg.
993 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
994 // which is returned in RAX / RDX.
995 if (Subtarget->is64Bit()) {
996 MVT ValVT = ValToCopy.getValueType();
997 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
998 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
999 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1000 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1004 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1005 Flag = Chain.getValue(1);
1008 // The x86-64 ABI for returning structs by value requires that we copy
1009 // the sret argument into %rax for the return. We saved the argument into
1010 // a virtual register in the entry block, so now we copy the value out
1012 if (Subtarget->is64Bit() &&
1013 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1014 MachineFunction &MF = DAG.getMachineFunction();
1015 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1016 unsigned Reg = FuncInfo->getSRetReturnReg();
1018 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1019 FuncInfo->setSRetReturnReg(Reg);
1021 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1023 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1024 Flag = Chain.getValue(1);
1027 RetOps[0] = Chain; // Update chain.
1029 // Add the flag if we have it.
1031 RetOps.push_back(Flag);
1033 return DAG.getNode(X86ISD::RET_FLAG, dl,
1034 MVT::Other, &RetOps[0], RetOps.size());
1038 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1039 /// appropriate copies out of appropriate physical registers. This assumes that
1040 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1041 /// being lowered. The returns a SDNode with the same number of values as the
1043 SDNode *X86TargetLowering::
1044 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1045 unsigned CallingConv, SelectionDAG &DAG) {
1047 DebugLoc dl = TheCall->getDebugLoc();
1048 // Assign locations to each value returned by this call.
1049 SmallVector<CCValAssign, 16> RVLocs;
1050 bool isVarArg = TheCall->isVarArg();
1051 bool Is64Bit = Subtarget->is64Bit();
1052 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1053 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1055 SmallVector<SDValue, 8> ResultVals;
1057 // Copy all of the result registers out of their specified physreg.
1058 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1059 CCValAssign &VA = RVLocs[i];
1060 MVT CopyVT = VA.getValVT();
1062 // If this is x86-64, and we disabled SSE, we can't return FP values
1063 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1064 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1065 cerr << "SSE register return with SSE disabled\n";
1069 // If this is a call to a function that returns an fp value on the floating
1070 // point stack, but where we prefer to use the value in xmm registers, copy
1071 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1072 if ((VA.getLocReg() == X86::ST0 ||
1073 VA.getLocReg() == X86::ST1) &&
1074 isScalarFPTypeInSSEReg(VA.getValVT())) {
1079 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1080 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1081 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1082 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1083 MVT::v2i64, InFlag).getValue(1);
1084 Val = Chain.getValue(0);
1085 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1086 Val, DAG.getConstant(0, MVT::i64));
1088 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1089 MVT::i64, InFlag).getValue(1);
1090 Val = Chain.getValue(0);
1092 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1094 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1095 CopyVT, InFlag).getValue(1);
1096 Val = Chain.getValue(0);
1098 InFlag = Chain.getValue(2);
1100 if (CopyVT != VA.getValVT()) {
1101 // Round the F80 the right size, which also moves to the appropriate xmm
1103 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1104 // This truncation won't change the value.
1105 DAG.getIntPtrConstant(1));
1108 ResultVals.push_back(Val);
1111 // Merge everything together with a MERGE_VALUES node.
1112 ResultVals.push_back(Chain);
1113 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1114 &ResultVals[0], ResultVals.size()).getNode();
1118 //===----------------------------------------------------------------------===//
1119 // C & StdCall & Fast Calling Convention implementation
1120 //===----------------------------------------------------------------------===//
1121 // StdCall calling convention seems to be standard for many Windows' API
1122 // routines and around. It differs from C calling convention just a little:
1123 // callee should clean up the stack, not caller. Symbols should be also
1124 // decorated in some fancy way :) It doesn't support any vector arguments.
1125 // For info on fast calling convention see Fast Calling Convention (tail call)
1126 // implementation LowerX86_32FastCCCallTo.
1128 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1130 static bool CallIsStructReturn(CallSDNode *TheCall) {
1131 unsigned NumOps = TheCall->getNumArgs();
1135 return TheCall->getArgFlags(0).isSRet();
1138 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1139 /// return semantics.
1140 static bool ArgsAreStructReturn(SDValue Op) {
1141 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1145 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1148 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1149 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1151 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1155 switch (CallingConv) {
1158 case CallingConv::X86_StdCall:
1159 return !Subtarget->is64Bit();
1160 case CallingConv::X86_FastCall:
1161 return !Subtarget->is64Bit();
1162 case CallingConv::Fast:
1163 return PerformTailCallOpt;
1167 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1168 /// given CallingConvention value.
1169 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1170 if (Subtarget->is64Bit()) {
1171 if (Subtarget->isTargetWin64())
1172 return CC_X86_Win64_C;
1177 if (CC == CallingConv::X86_FastCall)
1178 return CC_X86_32_FastCall;
1179 else if (CC == CallingConv::Fast)
1180 return CC_X86_32_FastCC;
1185 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1186 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1188 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1189 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1190 if (CC == CallingConv::X86_FastCall)
1192 else if (CC == CallingConv::X86_StdCall)
1198 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1199 /// in a register before calling.
1200 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1201 return !IsTailCall && !Is64Bit &&
1202 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1203 Subtarget->isPICStyleGOT();
1206 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1207 /// address to be loaded in a register.
1209 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1210 return !Is64Bit && IsTailCall &&
1211 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1212 Subtarget->isPICStyleGOT();
1215 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1216 /// by "Src" to address "Dst" with size and alignment information specified by
1217 /// the specific parameter attribute. The copy will be passed as a byval
1218 /// function parameter.
1220 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1221 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1223 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1224 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1225 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1228 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1229 const CCValAssign &VA,
1230 MachineFrameInfo *MFI,
1232 SDValue Root, unsigned i) {
1233 // Create the nodes corresponding to a load from this parameter slot.
1234 ISD::ArgFlagsTy Flags =
1235 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1236 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1237 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1239 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1240 // changed with more analysis.
1241 // In case of tail call optimization mark all arguments mutable. Since they
1242 // could be overwritten by lowering of arguments in case of a tail call.
1243 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1244 VA.getLocMemOffset(), isImmutable);
1245 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1246 if (Flags.isByVal())
1248 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1249 PseudoSourceValue::getFixedStack(FI), 0);
1253 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1254 MachineFunction &MF = DAG.getMachineFunction();
1255 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1256 DebugLoc dl = Op.getDebugLoc();
1258 const Function* Fn = MF.getFunction();
1259 if (Fn->hasExternalLinkage() &&
1260 Subtarget->isTargetCygMing() &&
1261 Fn->getName() == "main")
1262 FuncInfo->setForceFramePointer(true);
1264 // Decorate the function name.
1265 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1267 MachineFrameInfo *MFI = MF.getFrameInfo();
1268 SDValue Root = Op.getOperand(0);
1269 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1270 unsigned CC = MF.getFunction()->getCallingConv();
1271 bool Is64Bit = Subtarget->is64Bit();
1272 bool IsWin64 = Subtarget->isTargetWin64();
1274 assert(!(isVarArg && CC == CallingConv::Fast) &&
1275 "Var args not supported with calling convention fastcc");
1277 // Assign locations to all of the incoming arguments.
1278 SmallVector<CCValAssign, 16> ArgLocs;
1279 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1280 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1282 SmallVector<SDValue, 8> ArgValues;
1283 unsigned LastVal = ~0U;
1284 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1285 CCValAssign &VA = ArgLocs[i];
1286 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1288 assert(VA.getValNo() != LastVal &&
1289 "Don't support value assigned to multiple locs yet");
1290 LastVal = VA.getValNo();
1292 if (VA.isRegLoc()) {
1293 MVT RegVT = VA.getLocVT();
1294 TargetRegisterClass *RC = NULL;
1295 if (RegVT == MVT::i32)
1296 RC = X86::GR32RegisterClass;
1297 else if (Is64Bit && RegVT == MVT::i64)
1298 RC = X86::GR64RegisterClass;
1299 else if (RegVT == MVT::f32)
1300 RC = X86::FR32RegisterClass;
1301 else if (RegVT == MVT::f64)
1302 RC = X86::FR64RegisterClass;
1303 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1304 RC = X86::VR128RegisterClass;
1305 else if (RegVT.isVector()) {
1306 assert(RegVT.getSizeInBits() == 64);
1308 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1310 // Darwin calling convention passes MMX values in either GPRs or
1311 // XMMs in x86-64. Other targets pass them in memory.
1312 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1313 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1316 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1321 assert(0 && "Unknown argument type!");
1324 unsigned Reg = DAG.getMachineFunction().addLiveIn(VA.getLocReg(), RC);
1325 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1327 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1328 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1330 if (VA.getLocInfo() == CCValAssign::SExt)
1331 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1332 DAG.getValueType(VA.getValVT()));
1333 else if (VA.getLocInfo() == CCValAssign::ZExt)
1334 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1335 DAG.getValueType(VA.getValVT()));
1337 if (VA.getLocInfo() != CCValAssign::Full)
1338 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1340 // Handle MMX values passed in GPRs.
1341 if (Is64Bit && RegVT != VA.getLocVT()) {
1342 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1343 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1344 else if (RC == X86::VR128RegisterClass) {
1345 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1346 ArgValue, DAG.getConstant(0, MVT::i64));
1347 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1351 ArgValues.push_back(ArgValue);
1353 assert(VA.isMemLoc());
1354 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1358 // The x86-64 ABI for returning structs by value requires that we copy
1359 // the sret argument into %rax for the return. Save the argument into
1360 // a virtual register so that we can access it from the return points.
1361 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1362 MachineFunction &MF = DAG.getMachineFunction();
1363 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1364 unsigned Reg = FuncInfo->getSRetReturnReg();
1366 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1367 FuncInfo->setSRetReturnReg(Reg);
1369 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1370 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1373 unsigned StackSize = CCInfo.getNextStackOffset();
1374 // align stack specially for tail calls
1375 if (PerformTailCallOpt && CC == CallingConv::Fast)
1376 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1378 // If the function takes variable number of arguments, make a frame index for
1379 // the start of the first vararg value... for expansion of llvm.va_start.
1381 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1382 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1385 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1387 // FIXME: We should really autogenerate these arrays
1388 static const unsigned GPR64ArgRegsWin64[] = {
1389 X86::RCX, X86::RDX, X86::R8, X86::R9
1391 static const unsigned XMMArgRegsWin64[] = {
1392 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1394 static const unsigned GPR64ArgRegs64Bit[] = {
1395 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1397 static const unsigned XMMArgRegs64Bit[] = {
1398 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1399 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1401 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1404 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1405 GPR64ArgRegs = GPR64ArgRegsWin64;
1406 XMMArgRegs = XMMArgRegsWin64;
1408 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1409 GPR64ArgRegs = GPR64ArgRegs64Bit;
1410 XMMArgRegs = XMMArgRegs64Bit;
1412 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1414 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1417 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1418 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1419 "SSE register cannot be used when SSE is disabled!");
1420 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1421 "SSE register cannot be used when SSE is disabled!");
1422 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1423 // Kernel mode asks for SSE to be disabled, so don't push them
1425 TotalNumXMMRegs = 0;
1427 // For X86-64, if there are vararg parameters that are passed via
1428 // registers, then we must store them to their spots on the stack so they
1429 // may be loaded by deferencing the result of va_next.
1430 VarArgsGPOffset = NumIntRegs * 8;
1431 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1432 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1433 TotalNumXMMRegs * 16, 16);
1435 // Store the integer parameter registers.
1436 SmallVector<SDValue, 8> MemOps;
1437 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1438 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1439 DAG.getIntPtrConstant(VarArgsGPOffset));
1440 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1441 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1442 X86::GR64RegisterClass);
1443 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1445 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1446 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1447 MemOps.push_back(Store);
1448 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1449 DAG.getIntPtrConstant(8));
1452 // Now store the XMM (fp + vector) parameter registers.
1453 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1454 DAG.getIntPtrConstant(VarArgsFPOffset));
1455 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1456 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1457 X86::VR128RegisterClass);
1458 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1460 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1461 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1462 MemOps.push_back(Store);
1463 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1464 DAG.getIntPtrConstant(16));
1466 if (!MemOps.empty())
1467 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1468 &MemOps[0], MemOps.size());
1472 ArgValues.push_back(Root);
1474 // Some CCs need callee pop.
1475 if (IsCalleePop(isVarArg, CC)) {
1476 BytesToPopOnReturn = StackSize; // Callee pops everything.
1477 BytesCallerReserves = 0;
1479 BytesToPopOnReturn = 0; // Callee pops nothing.
1480 // If this is an sret function, the return should pop the hidden pointer.
1481 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1482 BytesToPopOnReturn = 4;
1483 BytesCallerReserves = StackSize;
1487 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1488 if (CC == CallingConv::X86_FastCall)
1489 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1492 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1494 // Return the new list of results.
1495 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1496 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1500 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1501 const SDValue &StackPtr,
1502 const CCValAssign &VA,
1504 SDValue Arg, ISD::ArgFlagsTy Flags) {
1505 DebugLoc dl = TheCall->getDebugLoc();
1506 unsigned LocMemOffset = VA.getLocMemOffset();
1507 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1508 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1509 if (Flags.isByVal()) {
1510 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1512 return DAG.getStore(Chain, dl, Arg, PtrOff,
1513 PseudoSourceValue::getStack(), LocMemOffset);
1516 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1517 /// optimization is performed and it is required.
1519 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1520 SDValue &OutRetAddr,
1526 if (!IsTailCall || FPDiff==0) return Chain;
1528 // Adjust the Return address stack slot.
1529 MVT VT = getPointerTy();
1530 OutRetAddr = getReturnAddressFrameIndex(DAG);
1532 // Load the "old" Return address.
1533 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1534 return SDValue(OutRetAddr.getNode(), 1);
1537 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1538 /// optimization is performed and it is required (FPDiff!=0).
1540 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1541 SDValue Chain, SDValue RetAddrFrIdx,
1542 bool Is64Bit, int FPDiff, DebugLoc dl) {
1543 // Store the return address to the appropriate stack slot.
1544 if (!FPDiff) return Chain;
1545 // Calculate the new stack slot for the return address.
1546 int SlotSize = Is64Bit ? 8 : 4;
1547 int NewReturnAddrFI =
1548 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1549 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1550 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1551 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1552 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1556 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1557 MachineFunction &MF = DAG.getMachineFunction();
1558 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1559 SDValue Chain = TheCall->getChain();
1560 unsigned CC = TheCall->getCallingConv();
1561 bool isVarArg = TheCall->isVarArg();
1562 bool IsTailCall = TheCall->isTailCall() &&
1563 CC == CallingConv::Fast && PerformTailCallOpt;
1564 SDValue Callee = TheCall->getCallee();
1565 bool Is64Bit = Subtarget->is64Bit();
1566 bool IsStructRet = CallIsStructReturn(TheCall);
1567 DebugLoc dl = TheCall->getDebugLoc();
1569 assert(!(isVarArg && CC == CallingConv::Fast) &&
1570 "Var args not supported with calling convention fastcc");
1572 // Analyze operands of the call, assigning locations to each operand.
1573 SmallVector<CCValAssign, 16> ArgLocs;
1574 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1575 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1577 // Get a count of how many bytes are to be pushed on the stack.
1578 unsigned NumBytes = CCInfo.getNextStackOffset();
1579 if (PerformTailCallOpt && CC == CallingConv::Fast)
1580 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1584 // Lower arguments at fp - stackoffset + fpdiff.
1585 unsigned NumBytesCallerPushed =
1586 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1587 FPDiff = NumBytesCallerPushed - NumBytes;
1589 // Set the delta of movement of the returnaddr stackslot.
1590 // But only set if delta is greater than previous delta.
1591 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1592 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1595 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1597 SDValue RetAddrFrIdx;
1598 // Load return adress for tail calls.
1599 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1602 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1603 SmallVector<SDValue, 8> MemOpChains;
1606 // Walk the register/memloc assignments, inserting copies/loads. In the case
1607 // of tail call optimization arguments are handle later.
1608 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1609 CCValAssign &VA = ArgLocs[i];
1610 SDValue Arg = TheCall->getArg(i);
1611 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1612 bool isByVal = Flags.isByVal();
1614 // Promote the value if needed.
1615 switch (VA.getLocInfo()) {
1616 default: assert(0 && "Unknown loc info!");
1617 case CCValAssign::Full: break;
1618 case CCValAssign::SExt:
1619 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1621 case CCValAssign::ZExt:
1622 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1624 case CCValAssign::AExt:
1625 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1629 if (VA.isRegLoc()) {
1631 MVT RegVT = VA.getLocVT();
1632 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1633 switch (VA.getLocReg()) {
1636 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1638 // Special case: passing MMX values in GPR registers.
1639 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1642 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1643 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1644 // Special case: passing MMX values in XMM registers.
1645 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1646 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1647 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1652 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1654 if (!IsTailCall || (IsTailCall && isByVal)) {
1655 assert(VA.isMemLoc());
1656 if (StackPtr.getNode() == 0)
1657 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1659 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1660 Chain, Arg, Flags));
1665 if (!MemOpChains.empty())
1666 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1667 &MemOpChains[0], MemOpChains.size());
1669 // Build a sequence of copy-to-reg nodes chained together with token chain
1670 // and flag operands which copy the outgoing args into registers.
1672 // Tail call byval lowering might overwrite argument registers so in case of
1673 // tail call optimization the copies to registers are lowered later.
1675 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1676 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1677 RegsToPass[i].second, InFlag);
1678 InFlag = Chain.getValue(1);
1681 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1683 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1684 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1685 DAG.getNode(X86ISD::GlobalBaseReg,
1686 DebugLoc::getUnknownLoc(),
1689 InFlag = Chain.getValue(1);
1691 // If we are tail calling and generating PIC/GOT style code load the address
1692 // of the callee into ecx. The value in ecx is used as target of the tail
1693 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1694 // calls on PIC/GOT architectures. Normally we would just put the address of
1695 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1696 // restored (since ebx is callee saved) before jumping to the target@PLT.
1697 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1698 // Note: The actual moving to ecx is done further down.
1699 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1700 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1701 !G->getGlobal()->hasProtectedVisibility())
1702 Callee = LowerGlobalAddress(Callee, DAG);
1703 else if (isa<ExternalSymbolSDNode>(Callee))
1704 Callee = LowerExternalSymbol(Callee,DAG);
1707 if (Is64Bit && isVarArg) {
1708 // From AMD64 ABI document:
1709 // For calls that may call functions that use varargs or stdargs
1710 // (prototype-less calls or calls to functions containing ellipsis (...) in
1711 // the declaration) %al is used as hidden argument to specify the number
1712 // of SSE registers used. The contents of %al do not need to match exactly
1713 // the number of registers, but must be an ubound on the number of SSE
1714 // registers used and is in the range 0 - 8 inclusive.
1716 // FIXME: Verify this on Win64
1717 // Count the number of XMM registers allocated.
1718 static const unsigned XMMArgRegs[] = {
1719 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1720 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1722 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1723 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1724 && "SSE registers cannot be used when SSE is disabled");
1726 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1727 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1728 InFlag = Chain.getValue(1);
1732 // For tail calls lower the arguments to the 'real' stack slot.
1734 SmallVector<SDValue, 8> MemOpChains2;
1737 // Do not flag preceeding copytoreg stuff together with the following stuff.
1739 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1740 CCValAssign &VA = ArgLocs[i];
1741 if (!VA.isRegLoc()) {
1742 assert(VA.isMemLoc());
1743 SDValue Arg = TheCall->getArg(i);
1744 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1745 // Create frame index.
1746 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1747 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1748 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1749 FIN = DAG.getFrameIndex(FI, getPointerTy());
1751 if (Flags.isByVal()) {
1752 // Copy relative to framepointer.
1753 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1754 if (StackPtr.getNode() == 0)
1755 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1757 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1759 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1762 // Store relative to framepointer.
1763 MemOpChains2.push_back(
1764 DAG.getStore(Chain, dl, Arg, FIN,
1765 PseudoSourceValue::getFixedStack(FI), 0));
1770 if (!MemOpChains2.empty())
1771 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1772 &MemOpChains2[0], MemOpChains2.size());
1774 // Copy arguments to their registers.
1775 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1776 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1777 RegsToPass[i].second, InFlag);
1778 InFlag = Chain.getValue(1);
1782 // Store the return address to the appropriate stack slot.
1783 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1787 // If the callee is a GlobalAddress node (quite common, every direct call is)
1788 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1789 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1790 // We should use extra load for direct calls to dllimported functions in
1792 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1793 getTargetMachine(), true))
1794 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1796 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1797 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1798 } else if (IsTailCall) {
1799 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1801 Chain = DAG.getCopyToReg(Chain, dl,
1802 DAG.getRegister(Opc, getPointerTy()),
1804 Callee = DAG.getRegister(Opc, getPointerTy());
1805 // Add register as live out.
1806 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1809 // Returns a chain & a flag for retval copy to use.
1810 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1811 SmallVector<SDValue, 8> Ops;
1814 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1815 DAG.getIntPtrConstant(0, true), InFlag);
1816 InFlag = Chain.getValue(1);
1818 // Returns a chain & a flag for retval copy to use.
1819 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1823 Ops.push_back(Chain);
1824 Ops.push_back(Callee);
1827 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1829 // Add argument registers to the end of the list so that they are known live
1831 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1832 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1833 RegsToPass[i].second.getValueType()));
1835 // Add an implicit use GOT pointer in EBX.
1836 if (!IsTailCall && !Is64Bit &&
1837 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1838 Subtarget->isPICStyleGOT())
1839 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1841 // Add an implicit use of AL for x86 vararg functions.
1842 if (Is64Bit && isVarArg)
1843 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1845 if (InFlag.getNode())
1846 Ops.push_back(InFlag);
1849 assert(InFlag.getNode() &&
1850 "Flag must be set. Depend on flag being set in LowerRET");
1851 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1852 TheCall->getVTList(), &Ops[0], Ops.size());
1854 return SDValue(Chain.getNode(), Op.getResNo());
1857 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1858 InFlag = Chain.getValue(1);
1860 // Create the CALLSEQ_END node.
1861 unsigned NumBytesForCalleeToPush;
1862 if (IsCalleePop(isVarArg, CC))
1863 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1864 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1865 // If this is is a call to a struct-return function, the callee
1866 // pops the hidden struct pointer, so we have to push it back.
1867 // This is common for Darwin/X86, Linux & Mingw32 targets.
1868 NumBytesForCalleeToPush = 4;
1870 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1872 // Returns a flag for retval copy to use.
1873 Chain = DAG.getCALLSEQ_END(Chain,
1874 DAG.getIntPtrConstant(NumBytes, true),
1875 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1878 InFlag = Chain.getValue(1);
1880 // Handle result values, copying them out of physregs into vregs that we
1882 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1887 //===----------------------------------------------------------------------===//
1888 // Fast Calling Convention (tail call) implementation
1889 //===----------------------------------------------------------------------===//
1891 // Like std call, callee cleans arguments, convention except that ECX is
1892 // reserved for storing the tail called function address. Only 2 registers are
1893 // free for argument passing (inreg). Tail call optimization is performed
1895 // * tailcallopt is enabled
1896 // * caller/callee are fastcc
1897 // On X86_64 architecture with GOT-style position independent code only local
1898 // (within module) calls are supported at the moment.
1899 // To keep the stack aligned according to platform abi the function
1900 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1901 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1902 // If a tail called function callee has more arguments than the caller the
1903 // caller needs to make sure that there is room to move the RETADDR to. This is
1904 // achieved by reserving an area the size of the argument delta right after the
1905 // original REtADDR, but before the saved framepointer or the spilled registers
1906 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1918 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1919 /// for a 16 byte align requirement.
1920 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1921 SelectionDAG& DAG) {
1922 MachineFunction &MF = DAG.getMachineFunction();
1923 const TargetMachine &TM = MF.getTarget();
1924 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1925 unsigned StackAlignment = TFI.getStackAlignment();
1926 uint64_t AlignMask = StackAlignment - 1;
1927 int64_t Offset = StackSize;
1928 uint64_t SlotSize = TD->getPointerSize();
1929 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1930 // Number smaller than 12 so just add the difference.
1931 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1933 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1934 Offset = ((~AlignMask) & Offset) + StackAlignment +
1935 (StackAlignment-SlotSize);
1940 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1941 /// following the call is a return. A function is eligible if caller/callee
1942 /// calling conventions match, currently only fastcc supports tail calls, and
1943 /// the function CALL is immediatly followed by a RET.
1944 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1946 SelectionDAG& DAG) const {
1947 if (!PerformTailCallOpt)
1950 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1951 MachineFunction &MF = DAG.getMachineFunction();
1952 unsigned CallerCC = MF.getFunction()->getCallingConv();
1953 unsigned CalleeCC= TheCall->getCallingConv();
1954 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1955 SDValue Callee = TheCall->getCallee();
1956 // On x86/32Bit PIC/GOT tail calls are supported.
1957 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1958 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1961 // Can only do local tail calls (in same module, hidden or protected) on
1962 // x86_64 PIC/GOT at the moment.
1963 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1964 return G->getGlobal()->hasHiddenVisibility()
1965 || G->getGlobal()->hasProtectedVisibility();
1973 X86TargetLowering::createFastISel(MachineFunction &mf,
1974 MachineModuleInfo *mmo,
1976 DenseMap<const Value *, unsigned> &vm,
1977 DenseMap<const BasicBlock *,
1978 MachineBasicBlock *> &bm,
1979 DenseMap<const AllocaInst *, int> &am
1981 , SmallSet<Instruction*, 8> &cil
1984 return X86::createFastISel(mf, mmo, dw, vm, bm, am
1992 //===----------------------------------------------------------------------===//
1993 // Other Lowering Hooks
1994 //===----------------------------------------------------------------------===//
1997 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1998 MachineFunction &MF = DAG.getMachineFunction();
1999 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2000 int ReturnAddrIndex = FuncInfo->getRAIndex();
2002 if (ReturnAddrIndex == 0) {
2003 // Set up a frame object for the return address.
2004 uint64_t SlotSize = TD->getPointerSize();
2005 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2006 FuncInfo->setRAIndex(ReturnAddrIndex);
2009 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2013 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2014 /// specific condition code, returning the condition code and the LHS/RHS of the
2015 /// comparison to make.
2016 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2017 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2019 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2020 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2021 // X > -1 -> X == 0, jump !sign.
2022 RHS = DAG.getConstant(0, RHS.getValueType());
2023 return X86::COND_NS;
2024 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2025 // X < 0 -> X == 0, jump on sign.
2027 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2029 RHS = DAG.getConstant(0, RHS.getValueType());
2030 return X86::COND_LE;
2034 switch (SetCCOpcode) {
2035 default: assert(0 && "Invalid integer condition!");
2036 case ISD::SETEQ: return X86::COND_E;
2037 case ISD::SETGT: return X86::COND_G;
2038 case ISD::SETGE: return X86::COND_GE;
2039 case ISD::SETLT: return X86::COND_L;
2040 case ISD::SETLE: return X86::COND_LE;
2041 case ISD::SETNE: return X86::COND_NE;
2042 case ISD::SETULT: return X86::COND_B;
2043 case ISD::SETUGT: return X86::COND_A;
2044 case ISD::SETULE: return X86::COND_BE;
2045 case ISD::SETUGE: return X86::COND_AE;
2049 // First determine if it is required or is profitable to flip the operands.
2051 // If LHS is a foldable load, but RHS is not, flip the condition.
2052 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2053 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2054 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2055 std::swap(LHS, RHS);
2058 switch (SetCCOpcode) {
2064 std::swap(LHS, RHS);
2068 // On a floating point condition, the flags are set as follows:
2070 // 0 | 0 | 0 | X > Y
2071 // 0 | 0 | 1 | X < Y
2072 // 1 | 0 | 0 | X == Y
2073 // 1 | 1 | 1 | unordered
2074 switch (SetCCOpcode) {
2075 default: assert(0 && "Condcode should be pre-legalized away");
2077 case ISD::SETEQ: return X86::COND_E;
2078 case ISD::SETOLT: // flipped
2080 case ISD::SETGT: return X86::COND_A;
2081 case ISD::SETOLE: // flipped
2083 case ISD::SETGE: return X86::COND_AE;
2084 case ISD::SETUGT: // flipped
2086 case ISD::SETLT: return X86::COND_B;
2087 case ISD::SETUGE: // flipped
2089 case ISD::SETLE: return X86::COND_BE;
2091 case ISD::SETNE: return X86::COND_NE;
2092 case ISD::SETUO: return X86::COND_P;
2093 case ISD::SETO: return X86::COND_NP;
2097 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2098 /// code. Current x86 isa includes the following FP cmov instructions:
2099 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2100 static bool hasFPCMov(unsigned X86CC) {
2116 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2117 /// the specified range (L, H].
2118 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2119 return (Val < 0) || (Val >= Low && Val < Hi);
2122 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2123 /// specified value.
2124 static bool isUndefOrEqual(int Val, int CmpVal) {
2125 if (Val < 0 || Val == CmpVal)
2130 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2131 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2132 /// the second operand.
2133 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2134 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2135 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2136 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2137 return (Mask[0] < 2 && Mask[1] < 2);
2141 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2142 SmallVector<int, 8> M;
2144 return ::isPSHUFDMask(M, N->getValueType(0));
2147 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2148 /// is suitable for input to PSHUFHW.
2149 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2150 if (VT != MVT::v8i16)
2153 // Lower quadword copied in order or undef.
2154 for (int i = 0; i != 4; ++i)
2155 if (Mask[i] >= 0 && Mask[i] != i)
2158 // Upper quadword shuffled.
2159 for (int i = 4; i != 8; ++i)
2160 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2166 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2167 SmallVector<int, 8> M;
2169 return ::isPSHUFHWMask(M, N->getValueType(0));
2172 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2173 /// is suitable for input to PSHUFLW.
2174 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2175 if (VT != MVT::v8i16)
2178 // Upper quadword copied in order.
2179 for (int i = 4; i != 8; ++i)
2180 if (Mask[i] >= 0 && Mask[i] != i)
2183 // Lower quadword shuffled.
2184 for (int i = 0; i != 4; ++i)
2191 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2192 SmallVector<int, 8> M;
2194 return ::isPSHUFLWMask(M, N->getValueType(0));
2197 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2198 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2199 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2200 int NumElems = VT.getVectorNumElements();
2201 if (NumElems != 2 && NumElems != 4)
2204 int Half = NumElems / 2;
2205 for (int i = 0; i < Half; ++i)
2206 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2208 for (int i = Half; i < NumElems; ++i)
2209 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2215 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2216 SmallVector<int, 8> M;
2218 return ::isSHUFPMask(M, N->getValueType(0));
2221 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2222 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2223 /// half elements to come from vector 1 (which would equal the dest.) and
2224 /// the upper half to come from vector 2.
2225 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2226 int NumElems = VT.getVectorNumElements();
2228 if (NumElems != 2 && NumElems != 4)
2231 int Half = NumElems / 2;
2232 for (int i = 0; i < Half; ++i)
2233 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2235 for (int i = Half; i < NumElems; ++i)
2236 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2241 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2242 SmallVector<int, 8> M;
2244 return isCommutedSHUFPMask(M, N->getValueType(0));
2247 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2248 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2249 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2250 if (N->getValueType(0).getVectorNumElements() != 4)
2253 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2254 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2255 isUndefOrEqual(N->getMaskElt(1), 7) &&
2256 isUndefOrEqual(N->getMaskElt(2), 2) &&
2257 isUndefOrEqual(N->getMaskElt(3), 3);
2260 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2261 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2262 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2263 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2265 if (NumElems != 2 && NumElems != 4)
2268 for (unsigned i = 0; i < NumElems/2; ++i)
2269 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2272 for (unsigned i = NumElems/2; i < NumElems; ++i)
2273 if (!isUndefOrEqual(N->getMaskElt(i), i))
2279 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2280 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2282 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2283 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2285 if (NumElems != 2 && NumElems != 4)
2288 for (unsigned i = 0; i < NumElems/2; ++i)
2289 if (!isUndefOrEqual(N->getMaskElt(i), i))
2292 for (unsigned i = 0; i < NumElems/2; ++i)
2293 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2299 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2300 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2302 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2303 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2308 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2309 isUndefOrEqual(N->getMaskElt(1), 3) &&
2310 isUndefOrEqual(N->getMaskElt(2), 2) &&
2311 isUndefOrEqual(N->getMaskElt(3), 3);
2314 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2315 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2316 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2317 bool V2IsSplat = false) {
2318 int NumElts = VT.getVectorNumElements();
2319 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2322 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2324 int BitI1 = Mask[i+1];
2325 if (!isUndefOrEqual(BitI, j))
2328 if (!isUndefOrEqual(BitI1, NumElts))
2331 if (!isUndefOrEqual(BitI1, j + NumElts))
2338 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2339 SmallVector<int, 8> M;
2341 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2344 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2345 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2346 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, MVT VT,
2347 bool V2IsSplat = false) {
2348 int NumElts = VT.getVectorNumElements();
2349 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2352 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2354 int BitI1 = Mask[i+1];
2355 if (!isUndefOrEqual(BitI, j + NumElts/2))
2358 if (isUndefOrEqual(BitI1, NumElts))
2361 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2368 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2369 SmallVector<int, 8> M;
2371 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2374 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2375 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2377 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2378 int NumElems = VT.getVectorNumElements();
2379 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2382 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2384 int BitI1 = Mask[i+1];
2385 if (!isUndefOrEqual(BitI, j))
2387 if (!isUndefOrEqual(BitI1, j))
2393 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2394 SmallVector<int, 8> M;
2396 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2399 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2400 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2402 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2403 int NumElems = VT.getVectorNumElements();
2404 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2407 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2409 int BitI1 = Mask[i+1];
2410 if (!isUndefOrEqual(BitI, j))
2412 if (!isUndefOrEqual(BitI1, j))
2418 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2419 SmallVector<int, 8> M;
2421 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2424 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2425 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2426 /// MOVSD, and MOVD, i.e. setting the lowest element.
2427 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2428 if (VT.getVectorElementType().getSizeInBits() < 32)
2431 int NumElts = VT.getVectorNumElements();
2433 if (!isUndefOrEqual(Mask[0], NumElts))
2436 for (int i = 1; i < NumElts; ++i)
2437 if (!isUndefOrEqual(Mask[i], i))
2443 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2444 SmallVector<int, 8> M;
2446 return ::isMOVLMask(M, N->getValueType(0));
2449 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2450 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2451 /// element of vector 2 and the other elements to come from vector 1 in order.
2452 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2453 bool V2IsSplat = false, bool V2IsUndef = false) {
2454 int NumOps = VT.getVectorNumElements();
2455 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2458 if (!isUndefOrEqual(Mask[0], 0))
2461 for (int i = 1; i < NumOps; ++i)
2462 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2463 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2464 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2470 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2471 bool V2IsUndef = false) {
2472 SmallVector<int, 8> M;
2474 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2477 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2478 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2479 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2480 if (N->getValueType(0).getVectorNumElements() != 4)
2483 // Expect 1, 1, 3, 3
2484 for (unsigned i = 0; i < 2; ++i) {
2485 int Elt = N->getMaskElt(i);
2486 if (Elt >= 0 && Elt != 1)
2491 for (unsigned i = 2; i < 4; ++i) {
2492 int Elt = N->getMaskElt(i);
2493 if (Elt >= 0 && Elt != 3)
2498 // Don't use movshdup if it can be done with a shufps.
2499 // FIXME: verify that matching u, u, 3, 3 is what we want.
2503 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2504 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2505 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2506 if (N->getValueType(0).getVectorNumElements() != 4)
2509 // Expect 0, 0, 2, 2
2510 for (unsigned i = 0; i < 2; ++i)
2511 if (N->getMaskElt(i) > 0)
2515 for (unsigned i = 2; i < 4; ++i) {
2516 int Elt = N->getMaskElt(i);
2517 if (Elt >= 0 && Elt != 2)
2522 // Don't use movsldup if it can be done with a shufps.
2526 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2527 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2528 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2529 int e = N->getValueType(0).getVectorNumElements() / 2;
2531 for (int i = 0; i < e; ++i)
2532 if (!isUndefOrEqual(N->getMaskElt(i), i))
2534 for (int i = 0; i < e; ++i)
2535 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2540 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2541 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2543 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2544 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2545 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2547 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2549 for (int i = 0; i < NumOperands; ++i) {
2550 int Val = SVOp->getMaskElt(NumOperands-i-1);
2551 if (Val < 0) Val = 0;
2552 if (Val >= NumOperands) Val -= NumOperands;
2554 if (i != NumOperands - 1)
2560 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2561 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2563 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2564 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2566 // 8 nodes, but we only care about the last 4.
2567 for (unsigned i = 7; i >= 4; --i) {
2568 int Val = SVOp->getMaskElt(i);
2577 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2578 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2580 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2581 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2583 // 8 nodes, but we only care about the first 4.
2584 for (int i = 3; i >= 0; --i) {
2585 int Val = SVOp->getMaskElt(i);
2594 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2595 /// their permute mask.
2596 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2597 SelectionDAG &DAG) {
2598 MVT VT = SVOp->getValueType(0);
2599 unsigned NumElems = VT.getVectorNumElements();
2600 SmallVector<int, 8> MaskVec;
2602 for (unsigned i = 0; i != NumElems; ++i) {
2603 int idx = SVOp->getMaskElt(i);
2605 MaskVec.push_back(idx);
2606 else if (idx < (int)NumElems)
2607 MaskVec.push_back(idx + NumElems);
2609 MaskVec.push_back(idx - NumElems);
2611 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2612 SVOp->getOperand(0), &MaskVec[0]);
2615 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2616 /// the two vector operands have swapped position.
2617 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, MVT VT) {
2618 unsigned NumElems = VT.getVectorNumElements();
2619 for (unsigned i = 0; i != NumElems; ++i) {
2623 else if (idx < (int)NumElems)
2624 Mask[i] = idx + NumElems;
2626 Mask[i] = idx - NumElems;
2630 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2631 /// match movhlps. The lower half elements should come from upper half of
2632 /// V1 (and in order), and the upper half elements should come from the upper
2633 /// half of V2 (and in order).
2634 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2635 if (Op->getValueType(0).getVectorNumElements() != 4)
2637 for (unsigned i = 0, e = 2; i != e; ++i)
2638 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2640 for (unsigned i = 2; i != 4; ++i)
2641 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2646 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2647 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2649 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2650 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2652 N = N->getOperand(0).getNode();
2653 if (!ISD::isNON_EXTLoad(N))
2656 *LD = cast<LoadSDNode>(N);
2660 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2661 /// match movlp{s|d}. The lower half elements should come from lower half of
2662 /// V1 (and in order), and the upper half elements should come from the upper
2663 /// half of V2 (and in order). And since V1 will become the source of the
2664 /// MOVLP, it must be either a vector load or a scalar load to vector.
2665 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2666 ShuffleVectorSDNode *Op) {
2667 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2669 // Is V2 is a vector load, don't do this transformation. We will try to use
2670 // load folding shufps op.
2671 if (ISD::isNON_EXTLoad(V2))
2674 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2676 if (NumElems != 2 && NumElems != 4)
2678 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2679 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2681 for (unsigned i = NumElems/2; i != NumElems; ++i)
2682 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2687 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2689 static bool isSplatVector(SDNode *N) {
2690 if (N->getOpcode() != ISD::BUILD_VECTOR)
2693 SDValue SplatValue = N->getOperand(0);
2694 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2695 if (N->getOperand(i) != SplatValue)
2700 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2702 static inline bool isZeroNode(SDValue Elt) {
2703 return ((isa<ConstantSDNode>(Elt) &&
2704 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2705 (isa<ConstantFPSDNode>(Elt) &&
2706 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2709 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2710 /// to an zero vector.
2711 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2712 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2713 SDValue V1 = N->getOperand(0);
2714 SDValue V2 = N->getOperand(1);
2715 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2716 for (unsigned i = 0; i != NumElems; ++i) {
2717 int Idx = N->getMaskElt(i);
2718 if (Idx >= (int)NumElems) {
2719 unsigned Opc = V2.getOpcode();
2720 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2722 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V2.getOperand(Idx-NumElems)))
2724 } else if (Idx >= 0) {
2725 unsigned Opc = V1.getOpcode();
2726 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2728 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V1.getOperand(Idx)))
2735 /// getZeroVector - Returns a vector of specified type with all zero elements.
2737 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2739 assert(VT.isVector() && "Expected a vector type");
2741 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2742 // type. This ensures they get CSE'd.
2744 if (VT.getSizeInBits() == 64) { // MMX
2745 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2746 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2747 } else if (HasSSE2) { // SSE2
2748 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2749 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2751 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2752 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2754 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2757 /// getOnesVector - Returns a vector of specified type with all bits set.
2759 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2760 assert(VT.isVector() && "Expected a vector type");
2762 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2763 // type. This ensures they get CSE'd.
2764 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2766 if (VT.getSizeInBits() == 64) // MMX
2767 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2769 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2770 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2774 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2775 /// that point to V2 points to its first element.
2776 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2777 MVT VT = SVOp->getValueType(0);
2778 unsigned NumElems = VT.getVectorNumElements();
2780 bool Changed = false;
2781 SmallVector<int, 8> MaskVec;
2782 SVOp->getMask(MaskVec);
2784 for (unsigned i = 0; i != NumElems; ++i) {
2785 if (MaskVec[i] > (int)NumElems) {
2786 MaskVec[i] = NumElems;
2791 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2792 SVOp->getOperand(1), &MaskVec[0]);
2793 return SDValue(SVOp, 0);
2796 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2797 /// operation of specified width.
2798 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2800 unsigned NumElems = VT.getVectorNumElements();
2801 SmallVector<int, 8> Mask;
2802 Mask.push_back(NumElems);
2803 for (unsigned i = 1; i != NumElems; ++i)
2805 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2808 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
2809 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2811 unsigned NumElems = VT.getVectorNumElements();
2812 SmallVector<int, 8> Mask;
2813 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2815 Mask.push_back(i + NumElems);
2817 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2820 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
2821 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2823 unsigned NumElems = VT.getVectorNumElements();
2824 unsigned Half = NumElems/2;
2825 SmallVector<int, 8> Mask;
2826 for (unsigned i = 0; i != Half; ++i) {
2827 Mask.push_back(i + Half);
2828 Mask.push_back(i + NumElems + Half);
2830 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2833 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2834 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2836 if (SV->getValueType(0).getVectorNumElements() <= 4)
2837 return SDValue(SV, 0);
2839 MVT PVT = MVT::v4f32;
2840 MVT VT = SV->getValueType(0);
2841 DebugLoc dl = SV->getDebugLoc();
2842 SDValue V1 = SV->getOperand(0);
2843 int NumElems = VT.getVectorNumElements();
2844 int EltNo = SV->getSplatIndex();
2846 // unpack elements to the correct location
2847 while (NumElems > 4) {
2848 if (EltNo < NumElems/2) {
2849 V1 = getUnpackl(DAG, dl, VT, V1, V1);
2851 V1 = getUnpackh(DAG, dl, VT, V1, V1);
2852 EltNo -= NumElems/2;
2857 // Perform the splat.
2858 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
2859 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
2860 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
2861 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
2864 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2865 /// vector of zero or undef vector. This produces a shuffle where the low
2866 /// element of V2 is swizzled into the zero/undef vector, landing at element
2867 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2868 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2869 bool isZero, bool HasSSE2,
2870 SelectionDAG &DAG) {
2871 MVT VT = V2.getValueType();
2873 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
2874 unsigned NumElems = VT.getVectorNumElements();
2875 SmallVector<int, 16> MaskVec;
2876 for (unsigned i = 0; i != NumElems; ++i)
2877 // If this is the insertion idx, put the low elt of V2 here.
2878 MaskVec.push_back(i == Idx ? NumElems : i);
2879 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
2882 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
2883 /// a shuffle that is zero.
2885 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
2886 bool Low, SelectionDAG &DAG) {
2887 unsigned NumZeros = 0;
2888 for (int i = 0; i < NumElems; ++i) {
2889 unsigned Index = Low ? i : NumElems-i-1;
2890 int Idx = SVOp->getMaskElt(Index);
2895 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
2896 if (Elt.getNode() && isZeroNode(Elt))
2904 /// isVectorShift - Returns true if the shuffle can be implemented as a
2905 /// logical left or right shift of a vector.
2906 /// FIXME: split into pslldqi, psrldqi, palignr variants.
2907 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
2908 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
2909 int NumElems = SVOp->getValueType(0).getVectorNumElements();
2912 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
2915 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
2919 bool SeenV1 = false;
2920 bool SeenV2 = false;
2921 for (int i = NumZeros; i < NumElems; ++i) {
2922 int Val = isLeft ? (i - NumZeros) : i;
2923 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
2935 if (SeenV1 && SeenV2)
2938 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
2944 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2946 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
2947 unsigned NumNonZero, unsigned NumZero,
2948 SelectionDAG &DAG, TargetLowering &TLI) {
2952 DebugLoc dl = Op.getDebugLoc();
2955 for (unsigned i = 0; i < 16; ++i) {
2956 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2957 if (ThisIsNonZero && First) {
2959 V = getZeroVector(MVT::v8i16, true, DAG, dl);
2961 V = DAG.getUNDEF(MVT::v8i16);
2966 SDValue ThisElt(0, 0), LastElt(0, 0);
2967 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2968 if (LastIsNonZero) {
2969 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
2970 MVT::i16, Op.getOperand(i-1));
2972 if (ThisIsNonZero) {
2973 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
2974 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
2975 ThisElt, DAG.getConstant(8, MVT::i8));
2977 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
2981 if (ThisElt.getNode())
2982 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
2983 DAG.getIntPtrConstant(i/2));
2987 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
2990 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
2992 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
2993 unsigned NumNonZero, unsigned NumZero,
2994 SelectionDAG &DAG, TargetLowering &TLI) {
2998 DebugLoc dl = Op.getDebugLoc();
3001 for (unsigned i = 0; i < 8; ++i) {
3002 bool isNonZero = (NonZeros & (1 << i)) != 0;
3006 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3008 V = DAG.getUNDEF(MVT::v8i16);
3011 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3012 MVT::v8i16, V, Op.getOperand(i),
3013 DAG.getIntPtrConstant(i));
3020 /// getVShift - Return a vector logical shift node.
3022 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3023 unsigned NumBits, SelectionDAG &DAG,
3024 const TargetLowering &TLI, DebugLoc dl) {
3025 bool isMMX = VT.getSizeInBits() == 64;
3026 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3027 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3028 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3029 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3030 DAG.getNode(Opc, dl, ShVT, SrcOp,
3031 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3035 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3036 DebugLoc dl = Op.getDebugLoc();
3037 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3038 if (ISD::isBuildVectorAllZeros(Op.getNode())
3039 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3040 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3041 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3042 // eliminated on x86-32 hosts.
3043 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3046 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3047 return getOnesVector(Op.getValueType(), DAG, dl);
3048 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3051 MVT VT = Op.getValueType();
3052 MVT EVT = VT.getVectorElementType();
3053 unsigned EVTBits = EVT.getSizeInBits();
3055 unsigned NumElems = Op.getNumOperands();
3056 unsigned NumZero = 0;
3057 unsigned NumNonZero = 0;
3058 unsigned NonZeros = 0;
3059 bool IsAllConstants = true;
3060 SmallSet<SDValue, 8> Values;
3061 for (unsigned i = 0; i < NumElems; ++i) {
3062 SDValue Elt = Op.getOperand(i);
3063 if (Elt.getOpcode() == ISD::UNDEF)
3066 if (Elt.getOpcode() != ISD::Constant &&
3067 Elt.getOpcode() != ISD::ConstantFP)
3068 IsAllConstants = false;
3069 if (isZeroNode(Elt))
3072 NonZeros |= (1 << i);
3077 if (NumNonZero == 0) {
3078 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3079 return DAG.getUNDEF(VT);
3082 // Special case for single non-zero, non-undef, element.
3083 if (NumNonZero == 1) {
3084 unsigned Idx = CountTrailingZeros_32(NonZeros);
3085 SDValue Item = Op.getOperand(Idx);
3087 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3088 // the value are obviously zero, truncate the value to i32 and do the
3089 // insertion that way. Only do this if the value is non-constant or if the
3090 // value is a constant being inserted into element 0. It is cheaper to do
3091 // a constant pool load than it is to do a movd + shuffle.
3092 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3093 (!IsAllConstants || Idx == 0)) {
3094 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3095 // Handle MMX and SSE both.
3096 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3097 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3099 // Truncate the value (which may itself be a constant) to i32, and
3100 // convert it to a vector with movd (S2V+shuffle to zero extend).
3101 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3102 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3103 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3104 Subtarget->hasSSE2(), DAG);
3106 // Now we have our 32-bit value zero extended in the low element of
3107 // a vector. If Idx != 0, swizzle it into place.
3109 SmallVector<int, 4> Mask;
3110 Mask.push_back(Idx);
3111 for (unsigned i = 1; i != VecElts; ++i)
3113 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3114 DAG.getUNDEF(Item.getValueType()),
3117 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3121 // If we have a constant or non-constant insertion into the low element of
3122 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3123 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3124 // depending on what the source datatype is.
3127 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3128 } else if (EVT == MVT::i32 || EVT == MVT::f32 || EVT == MVT::f64 ||
3129 (EVT == MVT::i64 && Subtarget->is64Bit())) {
3130 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3131 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3132 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3134 } else if (EVT == MVT::i16 || EVT == MVT::i8) {
3135 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3136 MVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3137 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3138 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3139 Subtarget->hasSSE2(), DAG);
3140 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3144 // Is it a vector logical left shift?
3145 if (NumElems == 2 && Idx == 1 &&
3146 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3147 unsigned NumBits = VT.getSizeInBits();
3148 return getVShift(true, VT,
3149 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3150 VT, Op.getOperand(1)),
3151 NumBits/2, DAG, *this, dl);
3154 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3157 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3158 // is a non-constant being inserted into an element other than the low one,
3159 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3160 // movd/movss) to move this into the low element, then shuffle it into
3162 if (EVTBits == 32) {
3163 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3165 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3166 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3167 Subtarget->hasSSE2(), DAG);
3168 SmallVector<int, 8> MaskVec;
3169 for (unsigned i = 0; i < NumElems; i++)
3170 MaskVec.push_back(i == Idx ? 0 : 1);
3171 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3175 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3176 if (Values.size() == 1)
3179 // A vector full of immediates; various special cases are already
3180 // handled, so this is best done with a single constant-pool load.
3184 // Let legalizer expand 2-wide build_vectors.
3185 if (EVTBits == 64) {
3186 if (NumNonZero == 1) {
3187 // One half is zero or undef.
3188 unsigned Idx = CountTrailingZeros_32(NonZeros);
3189 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3190 Op.getOperand(Idx));
3191 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3192 Subtarget->hasSSE2(), DAG);
3197 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3198 if (EVTBits == 8 && NumElems == 16) {
3199 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3201 if (V.getNode()) return V;
3204 if (EVTBits == 16 && NumElems == 8) {
3205 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3207 if (V.getNode()) return V;
3210 // If element VT is == 32 bits, turn it into a number of shuffles.
3211 SmallVector<SDValue, 8> V;
3213 if (NumElems == 4 && NumZero > 0) {
3214 for (unsigned i = 0; i < 4; ++i) {
3215 bool isZero = !(NonZeros & (1 << i));
3217 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3219 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3222 for (unsigned i = 0; i < 2; ++i) {
3223 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3226 V[i] = V[i*2]; // Must be a zero vector.
3229 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3232 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3235 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3240 SmallVector<int, 8> MaskVec;
3241 bool Reverse = (NonZeros & 0x3) == 2;
3242 for (unsigned i = 0; i < 2; ++i)
3243 MaskVec.push_back(Reverse ? 1-i : i);
3244 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3245 for (unsigned i = 0; i < 2; ++i)
3246 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3247 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3250 if (Values.size() > 2) {
3251 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3252 // values to be inserted is equal to the number of elements, in which case
3253 // use the unpack code below in the hopes of matching the consecutive elts
3254 // load merge pattern for shuffles.
3255 // FIXME: We could probably just check that here directly.
3256 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3257 getSubtarget()->hasSSE41()) {
3258 V[0] = DAG.getUNDEF(VT);
3259 for (unsigned i = 0; i < NumElems; ++i)
3260 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3261 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3262 Op.getOperand(i), DAG.getIntPtrConstant(i));
3265 // Expand into a number of unpckl*.
3267 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3268 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3269 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3270 for (unsigned i = 0; i < NumElems; ++i)
3271 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3273 while (NumElems != 0) {
3274 for (unsigned i = 0; i < NumElems; ++i)
3275 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3284 // v8i16 shuffles - Prefer shuffles in the following order:
3285 // 1. [all] pshuflw, pshufhw, optional move
3286 // 2. [ssse3] 1 x pshufb
3287 // 3. [ssse3] 2 x pshufb + 1 x por
3288 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3290 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3291 SelectionDAG &DAG, X86TargetLowering &TLI) {
3292 SDValue V1 = SVOp->getOperand(0);
3293 SDValue V2 = SVOp->getOperand(1);
3294 DebugLoc dl = SVOp->getDebugLoc();
3295 SmallVector<int, 8> MaskVals;
3297 // Determine if more than 1 of the words in each of the low and high quadwords
3298 // of the result come from the same quadword of one of the two inputs. Undef
3299 // mask values count as coming from any quadword, for better codegen.
3300 SmallVector<unsigned, 4> LoQuad(4);
3301 SmallVector<unsigned, 4> HiQuad(4);
3302 BitVector InputQuads(4);
3303 for (unsigned i = 0; i < 8; ++i) {
3304 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3305 int EltIdx = SVOp->getMaskElt(i);
3306 MaskVals.push_back(EltIdx);
3315 InputQuads.set(EltIdx / 4);
3318 int BestLoQuad = -1;
3319 unsigned MaxQuad = 1;
3320 for (unsigned i = 0; i < 4; ++i) {
3321 if (LoQuad[i] > MaxQuad) {
3323 MaxQuad = LoQuad[i];
3327 int BestHiQuad = -1;
3329 for (unsigned i = 0; i < 4; ++i) {
3330 if (HiQuad[i] > MaxQuad) {
3332 MaxQuad = HiQuad[i];
3336 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3337 // of the two input vectors, shuffle them into one input vector so only a
3338 // single pshufb instruction is necessary. If There are more than 2 input
3339 // quads, disable the next transformation since it does not help SSSE3.
3340 bool V1Used = InputQuads[0] || InputQuads[1];
3341 bool V2Used = InputQuads[2] || InputQuads[3];
3342 if (TLI.getSubtarget()->hasSSSE3()) {
3343 if (InputQuads.count() == 2 && V1Used && V2Used) {
3344 BestLoQuad = InputQuads.find_first();
3345 BestHiQuad = InputQuads.find_next(BestLoQuad);
3347 if (InputQuads.count() > 2) {
3353 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3354 // the shuffle mask. If a quad is scored as -1, that means that it contains
3355 // words from all 4 input quadwords.
3357 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3358 SmallVector<int, 8> MaskV;
3359 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3360 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3361 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3362 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3363 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3364 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3366 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3367 // source words for the shuffle, to aid later transformations.
3368 bool AllWordsInNewV = true;
3369 bool InOrder[2] = { true, true };
3370 for (unsigned i = 0; i != 8; ++i) {
3371 int idx = MaskVals[i];
3373 InOrder[i/4] = false;
3374 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3376 AllWordsInNewV = false;
3380 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3381 if (AllWordsInNewV) {
3382 for (int i = 0; i != 8; ++i) {
3383 int idx = MaskVals[i];
3386 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3387 if ((idx != i) && idx < 4)
3389 if ((idx != i) && idx > 3)
3398 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3399 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3400 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3401 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3402 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3406 // If we have SSSE3, and all words of the result are from 1 input vector,
3407 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3408 // is present, fall back to case 4.
3409 if (TLI.getSubtarget()->hasSSSE3()) {
3410 SmallVector<SDValue,16> pshufbMask;
3412 // If we have elements from both input vectors, set the high bit of the
3413 // shuffle mask element to zero out elements that come from V2 in the V1
3414 // mask, and elements that come from V1 in the V2 mask, so that the two
3415 // results can be OR'd together.
3416 bool TwoInputs = V1Used && V2Used;
3417 for (unsigned i = 0; i != 8; ++i) {
3418 int EltIdx = MaskVals[i] * 2;
3419 if (TwoInputs && (EltIdx >= 16)) {
3420 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3421 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3424 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3425 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3427 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3428 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3429 DAG.getNode(ISD::BUILD_VECTOR, dl,
3430 MVT::v16i8, &pshufbMask[0], 16));
3432 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3434 // Calculate the shuffle mask for the second input, shuffle it, and
3435 // OR it with the first shuffled input.
3437 for (unsigned i = 0; i != 8; ++i) {
3438 int EltIdx = MaskVals[i] * 2;
3440 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3441 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3444 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3445 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3447 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3448 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3449 DAG.getNode(ISD::BUILD_VECTOR, dl,
3450 MVT::v16i8, &pshufbMask[0], 16));
3451 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3452 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3455 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3456 // and update MaskVals with new element order.
3457 BitVector InOrder(8);
3458 if (BestLoQuad >= 0) {
3459 SmallVector<int, 8> MaskV;
3460 for (int i = 0; i != 4; ++i) {
3461 int idx = MaskVals[i];
3463 MaskV.push_back(-1);
3465 } else if ((idx / 4) == BestLoQuad) {
3466 MaskV.push_back(idx & 3);
3469 MaskV.push_back(-1);
3472 for (unsigned i = 4; i != 8; ++i)
3474 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3478 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3479 // and update MaskVals with the new element order.
3480 if (BestHiQuad >= 0) {
3481 SmallVector<int, 8> MaskV;
3482 for (unsigned i = 0; i != 4; ++i)
3484 for (unsigned i = 4; i != 8; ++i) {
3485 int idx = MaskVals[i];
3487 MaskV.push_back(-1);
3489 } else if ((idx / 4) == BestHiQuad) {
3490 MaskV.push_back((idx & 3) + 4);
3493 MaskV.push_back(-1);
3496 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3500 // In case BestHi & BestLo were both -1, which means each quadword has a word
3501 // from each of the four input quadwords, calculate the InOrder bitvector now
3502 // before falling through to the insert/extract cleanup.
3503 if (BestLoQuad == -1 && BestHiQuad == -1) {
3505 for (int i = 0; i != 8; ++i)
3506 if (MaskVals[i] < 0 || MaskVals[i] == i)
3510 // The other elements are put in the right place using pextrw and pinsrw.
3511 for (unsigned i = 0; i != 8; ++i) {
3514 int EltIdx = MaskVals[i];
3517 SDValue ExtOp = (EltIdx < 8)
3518 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3519 DAG.getIntPtrConstant(EltIdx))
3520 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3521 DAG.getIntPtrConstant(EltIdx - 8));
3522 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3523 DAG.getIntPtrConstant(i));
3528 // v16i8 shuffles - Prefer shuffles in the following order:
3529 // 1. [ssse3] 1 x pshufb
3530 // 2. [ssse3] 2 x pshufb + 1 x por
3531 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3533 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3534 SelectionDAG &DAG, X86TargetLowering &TLI) {
3535 SDValue V1 = SVOp->getOperand(0);
3536 SDValue V2 = SVOp->getOperand(1);
3537 DebugLoc dl = SVOp->getDebugLoc();
3538 SmallVector<int, 16> MaskVals;
3539 SVOp->getMask(MaskVals);
3541 // If we have SSSE3, case 1 is generated when all result bytes come from
3542 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3543 // present, fall back to case 3.
3544 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3547 for (unsigned i = 0; i < 16; ++i) {
3548 int EltIdx = MaskVals[i];
3557 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3558 if (TLI.getSubtarget()->hasSSSE3()) {
3559 SmallVector<SDValue,16> pshufbMask;
3561 // If all result elements are from one input vector, then only translate
3562 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3564 // Otherwise, we have elements from both input vectors, and must zero out
3565 // elements that come from V2 in the first mask, and V1 in the second mask
3566 // so that we can OR them together.
3567 bool TwoInputs = !(V1Only || V2Only);
3568 for (unsigned i = 0; i != 16; ++i) {
3569 int EltIdx = MaskVals[i];
3570 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3571 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3574 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3576 // If all the elements are from V2, assign it to V1 and return after
3577 // building the first pshufb.
3580 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3581 DAG.getNode(ISD::BUILD_VECTOR, dl,
3582 MVT::v16i8, &pshufbMask[0], 16));
3586 // Calculate the shuffle mask for the second input, shuffle it, and
3587 // OR it with the first shuffled input.
3589 for (unsigned i = 0; i != 16; ++i) {
3590 int EltIdx = MaskVals[i];
3592 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3595 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3597 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3598 DAG.getNode(ISD::BUILD_VECTOR, dl,
3599 MVT::v16i8, &pshufbMask[0], 16));
3600 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3603 // No SSSE3 - Calculate in place words and then fix all out of place words
3604 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3605 // the 16 different words that comprise the two doublequadword input vectors.
3606 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3607 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3608 SDValue NewV = V2Only ? V2 : V1;
3609 for (int i = 0; i != 8; ++i) {
3610 int Elt0 = MaskVals[i*2];
3611 int Elt1 = MaskVals[i*2+1];
3613 // This word of the result is all undef, skip it.
3614 if (Elt0 < 0 && Elt1 < 0)
3617 // This word of the result is already in the correct place, skip it.
3618 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3620 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3623 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3624 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3627 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3628 // using a single extract together, load it and store it.
3629 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3630 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3631 DAG.getIntPtrConstant(Elt1 / 2));
3632 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3633 DAG.getIntPtrConstant(i));
3637 // If Elt1 is defined, extract it from the appropriate source. If the
3638 // source byte is not also odd, shift the extracted word left 8 bits
3639 // otherwise clear the bottom 8 bits if we need to do an or.
3641 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3642 DAG.getIntPtrConstant(Elt1 / 2));
3643 if ((Elt1 & 1) == 0)
3644 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3645 DAG.getConstant(8, TLI.getShiftAmountTy()));
3647 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3648 DAG.getConstant(0xFF00, MVT::i16));
3650 // If Elt0 is defined, extract it from the appropriate source. If the
3651 // source byte is not also even, shift the extracted word right 8 bits. If
3652 // Elt1 was also defined, OR the extracted values together before
3653 // inserting them in the result.
3655 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3656 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3657 if ((Elt0 & 1) != 0)
3658 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3659 DAG.getConstant(8, TLI.getShiftAmountTy()));
3661 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3662 DAG.getConstant(0x00FF, MVT::i16));
3663 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3666 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3667 DAG.getIntPtrConstant(i));
3669 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3672 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3673 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3674 /// done when every pair / quad of shuffle mask elements point to elements in
3675 /// the right sequence. e.g.
3676 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3678 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3680 TargetLowering &TLI, DebugLoc dl) {
3681 MVT VT = SVOp->getValueType(0);
3682 SDValue V1 = SVOp->getOperand(0);
3683 SDValue V2 = SVOp->getOperand(1);
3684 unsigned NumElems = VT.getVectorNumElements();
3685 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3686 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3687 MVT MaskEltVT = MaskVT.getVectorElementType();
3689 switch (VT.getSimpleVT()) {
3690 default: assert(false && "Unexpected!");
3691 case MVT::v4f32: NewVT = MVT::v2f64; break;
3692 case MVT::v4i32: NewVT = MVT::v2i64; break;
3693 case MVT::v8i16: NewVT = MVT::v4i32; break;
3694 case MVT::v16i8: NewVT = MVT::v4i32; break;
3697 if (NewWidth == 2) {
3703 int Scale = NumElems / NewWidth;
3704 SmallVector<int, 8> MaskVec;
3705 for (unsigned i = 0; i < NumElems; i += Scale) {
3707 for (int j = 0; j < Scale; ++j) {
3708 int EltIdx = SVOp->getMaskElt(i+j);
3712 StartIdx = EltIdx - (EltIdx % Scale);
3713 if (EltIdx != StartIdx + j)
3717 MaskVec.push_back(-1);
3719 MaskVec.push_back(StartIdx / Scale);
3722 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3723 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3724 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3727 /// getVZextMovL - Return a zero-extending vector move low node.
3729 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3730 SDValue SrcOp, SelectionDAG &DAG,
3731 const X86Subtarget *Subtarget, DebugLoc dl) {
3732 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3733 LoadSDNode *LD = NULL;
3734 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3735 LD = dyn_cast<LoadSDNode>(SrcOp);
3737 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3739 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3740 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3741 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3742 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3743 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3745 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3746 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3747 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3748 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3756 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3757 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3758 DAG.getNode(ISD::BIT_CONVERT, dl,
3762 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3765 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3766 SDValue V1 = SVOp->getOperand(0);
3767 SDValue V2 = SVOp->getOperand(1);
3768 DebugLoc dl = SVOp->getDebugLoc();
3769 MVT VT = SVOp->getValueType(0);
3771 SmallVector<std::pair<int, int>, 8> Locs;
3773 SmallVector<int, 8> Mask1(4U, -1);
3774 SmallVector<int, 8> PermMask;
3775 SVOp->getMask(PermMask);
3779 for (unsigned i = 0; i != 4; ++i) {
3780 int Idx = PermMask[i];
3782 Locs[i] = std::make_pair(-1, -1);
3784 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3786 Locs[i] = std::make_pair(0, NumLo);
3790 Locs[i] = std::make_pair(1, NumHi);
3792 Mask1[2+NumHi] = Idx;
3798 if (NumLo <= 2 && NumHi <= 2) {
3799 // If no more than two elements come from either vector. This can be
3800 // implemented with two shuffles. First shuffle gather the elements.
3801 // The second shuffle, which takes the first shuffle as both of its
3802 // vector operands, put the elements into the right order.
3803 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3805 SmallVector<int, 8> Mask2(4U, -1);
3807 for (unsigned i = 0; i != 4; ++i) {
3808 if (Locs[i].first == -1)
3811 unsigned Idx = (i < 2) ? 0 : 4;
3812 Idx += Locs[i].first * 2 + Locs[i].second;
3817 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
3818 } else if (NumLo == 3 || NumHi == 3) {
3819 // Otherwise, we must have three elements from one vector, call it X, and
3820 // one element from the other, call it Y. First, use a shufps to build an
3821 // intermediate vector with the one element from Y and the element from X
3822 // that will be in the same half in the final destination (the indexes don't
3823 // matter). Then, use a shufps to build the final vector, taking the half
3824 // containing the element from Y from the intermediate, and the other half
3827 // Normalize it so the 3 elements come from V1.
3828 CommuteVectorShuffleMask(PermMask, VT);
3832 // Find the element from V2.
3834 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3835 int Val = PermMask[HiIndex];
3842 Mask1[0] = PermMask[HiIndex];
3844 Mask1[2] = PermMask[HiIndex^1];
3846 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3849 Mask1[0] = PermMask[0];
3850 Mask1[1] = PermMask[1];
3851 Mask1[2] = HiIndex & 1 ? 6 : 4;
3852 Mask1[3] = HiIndex & 1 ? 4 : 6;
3853 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3855 Mask1[0] = HiIndex & 1 ? 2 : 0;
3856 Mask1[1] = HiIndex & 1 ? 0 : 2;
3857 Mask1[2] = PermMask[2];
3858 Mask1[3] = PermMask[3];
3863 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
3867 // Break it into (shuffle shuffle_hi, shuffle_lo).
3869 SmallVector<int,8> LoMask(4U, -1);
3870 SmallVector<int,8> HiMask(4U, -1);
3872 SmallVector<int,8> *MaskPtr = &LoMask;
3873 unsigned MaskIdx = 0;
3876 for (unsigned i = 0; i != 4; ++i) {
3883 int Idx = PermMask[i];
3885 Locs[i] = std::make_pair(-1, -1);
3886 } else if (Idx < 4) {
3887 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3888 (*MaskPtr)[LoIdx] = Idx;
3891 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3892 (*MaskPtr)[HiIdx] = Idx;
3897 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
3898 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
3899 SmallVector<int, 8> MaskOps;
3900 for (unsigned i = 0; i != 4; ++i) {
3901 if (Locs[i].first == -1) {
3902 MaskOps.push_back(-1);
3904 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3905 MaskOps.push_back(Idx);
3908 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
3912 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3913 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
3914 SDValue V1 = Op.getOperand(0);
3915 SDValue V2 = Op.getOperand(1);
3916 MVT VT = Op.getValueType();
3917 DebugLoc dl = Op.getDebugLoc();
3918 unsigned NumElems = VT.getVectorNumElements();
3919 bool isMMX = VT.getSizeInBits() == 64;
3920 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3921 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3922 bool V1IsSplat = false;
3923 bool V2IsSplat = false;
3925 if (isZeroShuffle(SVOp))
3926 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3928 // Promote splats to v4f32.
3929 if (SVOp->isSplat()) {
3930 if (isMMX || NumElems < 4)
3932 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
3935 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3937 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3938 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
3939 if (NewOp.getNode())
3940 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3941 LowerVECTOR_SHUFFLE(NewOp, DAG));
3942 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3943 // FIXME: Figure out a cleaner way to do this.
3944 // Try to make use of movq to zero out the top part.
3945 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
3946 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
3947 if (NewOp.getNode()) {
3948 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
3949 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
3950 DAG, Subtarget, dl);
3952 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
3953 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
3954 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
3955 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
3956 DAG, Subtarget, dl);
3960 if (X86::isPSHUFDMask(SVOp))
3963 // Check if this can be converted into a logical shift.
3964 bool isLeft = false;
3967 bool isShift = getSubtarget()->hasSSE2() &&
3968 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
3969 if (isShift && ShVal.hasOneUse()) {
3970 // If the shifted value has multiple uses, it may be cheaper to use
3971 // v_set0 + movlhps or movhlps, etc.
3972 MVT EVT = VT.getVectorElementType();
3973 ShAmt *= EVT.getSizeInBits();
3974 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
3977 if (X86::isMOVLMask(SVOp)) {
3980 if (ISD::isBuildVectorAllZeros(V1.getNode()))
3981 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
3986 // FIXME: fold these into legal mask.
3987 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
3988 X86::isMOVSLDUPMask(SVOp) ||
3989 X86::isMOVHLPSMask(SVOp) ||
3990 X86::isMOVHPMask(SVOp) ||
3991 X86::isMOVLPMask(SVOp)))
3994 if (ShouldXformToMOVHLPS(SVOp) ||
3995 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
3996 return CommuteVectorShuffle(SVOp, DAG);
3999 // No better options. Use a vshl / vsrl.
4000 MVT EVT = VT.getVectorElementType();
4001 ShAmt *= EVT.getSizeInBits();
4002 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4005 bool Commuted = false;
4006 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4007 // 1,1,1,1 -> v8i16 though.
4008 V1IsSplat = isSplatVector(V1.getNode());
4009 V2IsSplat = isSplatVector(V2.getNode());
4011 // Canonicalize the splat or undef, if present, to be on the RHS.
4012 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4013 Op = CommuteVectorShuffle(SVOp, DAG);
4014 SVOp = cast<ShuffleVectorSDNode>(Op);
4015 V1 = SVOp->getOperand(0);
4016 V2 = SVOp->getOperand(1);
4017 std::swap(V1IsSplat, V2IsSplat);
4018 std::swap(V1IsUndef, V2IsUndef);
4022 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4023 // Shuffling low element of v1 into undef, just return v1.
4026 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4027 // the instruction selector will not match, so get a canonical MOVL with
4028 // swapped operands to undo the commute.
4029 return getMOVL(DAG, dl, VT, V2, V1);
4032 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4033 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4034 X86::isUNPCKLMask(SVOp) ||
4035 X86::isUNPCKHMask(SVOp))
4039 // Normalize mask so all entries that point to V2 points to its first
4040 // element then try to match unpck{h|l} again. If match, return a
4041 // new vector_shuffle with the corrected mask.
4042 SDValue NewMask = NormalizeMask(SVOp, DAG);
4043 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4044 if (NSVOp != SVOp) {
4045 if (X86::isUNPCKLMask(NSVOp, true)) {
4047 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4054 // Commute is back and try unpck* again.
4055 // FIXME: this seems wrong.
4056 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4057 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4058 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4059 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4060 X86::isUNPCKLMask(NewSVOp) ||
4061 X86::isUNPCKHMask(NewSVOp))
4065 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4067 // Normalize the node to match x86 shuffle ops if needed
4068 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4069 return CommuteVectorShuffle(SVOp, DAG);
4071 // Check for legal shuffle and return?
4072 SmallVector<int, 16> PermMask;
4073 SVOp->getMask(PermMask);
4074 if (isShuffleMaskLegal(PermMask, VT))
4077 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4078 if (VT == MVT::v8i16) {
4079 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4080 if (NewOp.getNode())
4084 if (VT == MVT::v16i8) {
4085 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4086 if (NewOp.getNode())
4090 // Handle all 4 wide cases with a number of shuffles except for MMX.
4091 if (NumElems == 4 && !isMMX)
4092 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4098 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4099 SelectionDAG &DAG) {
4100 MVT VT = Op.getValueType();
4101 DebugLoc dl = Op.getDebugLoc();
4102 if (VT.getSizeInBits() == 8) {
4103 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4104 Op.getOperand(0), Op.getOperand(1));
4105 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4106 DAG.getValueType(VT));
4107 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4108 } else if (VT.getSizeInBits() == 16) {
4109 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4110 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4112 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4113 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4114 DAG.getNode(ISD::BIT_CONVERT, dl,
4118 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4119 Op.getOperand(0), Op.getOperand(1));
4120 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4121 DAG.getValueType(VT));
4122 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4123 } else if (VT == MVT::f32) {
4124 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4125 // the result back to FR32 register. It's only worth matching if the
4126 // result has a single use which is a store or a bitcast to i32. And in
4127 // the case of a store, it's not worth it if the index is a constant 0,
4128 // because a MOVSSmr can be used instead, which is smaller and faster.
4129 if (!Op.hasOneUse())
4131 SDNode *User = *Op.getNode()->use_begin();
4132 if ((User->getOpcode() != ISD::STORE ||
4133 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4134 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4135 (User->getOpcode() != ISD::BIT_CONVERT ||
4136 User->getValueType(0) != MVT::i32))
4138 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4139 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4142 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4143 } else if (VT == MVT::i32) {
4144 // ExtractPS works with constant index.
4145 if (isa<ConstantSDNode>(Op.getOperand(1)))
4153 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4154 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4157 if (Subtarget->hasSSE41()) {
4158 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4163 MVT VT = Op.getValueType();
4164 DebugLoc dl = Op.getDebugLoc();
4165 // TODO: handle v16i8.
4166 if (VT.getSizeInBits() == 16) {
4167 SDValue Vec = Op.getOperand(0);
4168 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4170 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4171 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4172 DAG.getNode(ISD::BIT_CONVERT, dl,
4175 // Transform it so it match pextrw which produces a 32-bit result.
4176 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4177 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4178 Op.getOperand(0), Op.getOperand(1));
4179 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4180 DAG.getValueType(VT));
4181 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4182 } else if (VT.getSizeInBits() == 32) {
4183 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4187 // SHUFPS the element to the lowest double word, then movss.
4188 int Mask[4] = { Idx, -1, -1, -1 };
4189 MVT VVT = Op.getOperand(0).getValueType();
4190 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4191 DAG.getUNDEF(VVT), Mask);
4192 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4193 DAG.getIntPtrConstant(0));
4194 } else if (VT.getSizeInBits() == 64) {
4195 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4196 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4197 // to match extract_elt for f64.
4198 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4202 // UNPCKHPD the element to the lowest double word, then movsd.
4203 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4204 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4205 int Mask[2] = { 1, -1 };
4206 MVT VVT = Op.getOperand(0).getValueType();
4207 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4208 DAG.getUNDEF(VVT), Mask);
4209 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4210 DAG.getIntPtrConstant(0));
4217 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4218 MVT VT = Op.getValueType();
4219 MVT EVT = VT.getVectorElementType();
4220 DebugLoc dl = Op.getDebugLoc();
4222 SDValue N0 = Op.getOperand(0);
4223 SDValue N1 = Op.getOperand(1);
4224 SDValue N2 = Op.getOperand(2);
4226 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4227 isa<ConstantSDNode>(N2)) {
4228 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4230 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4232 if (N1.getValueType() != MVT::i32)
4233 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4234 if (N2.getValueType() != MVT::i32)
4235 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4236 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4237 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4238 // Bits [7:6] of the constant are the source select. This will always be
4239 // zero here. The DAG Combiner may combine an extract_elt index into these
4240 // bits. For example (insert (extract, 3), 2) could be matched by putting
4241 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4242 // Bits [5:4] of the constant are the destination select. This is the
4243 // value of the incoming immediate.
4244 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4245 // combine either bitwise AND or insert of float 0.0 to set these bits.
4246 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4247 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4248 } else if (EVT == MVT::i32) {
4249 // InsertPS works with constant index.
4250 if (isa<ConstantSDNode>(N2))
4257 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4258 MVT VT = Op.getValueType();
4259 MVT EVT = VT.getVectorElementType();
4261 if (Subtarget->hasSSE41())
4262 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4267 DebugLoc dl = Op.getDebugLoc();
4268 SDValue N0 = Op.getOperand(0);
4269 SDValue N1 = Op.getOperand(1);
4270 SDValue N2 = Op.getOperand(2);
4272 if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4273 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4274 // as its second argument.
4275 if (N1.getValueType() != MVT::i32)
4276 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4277 if (N2.getValueType() != MVT::i32)
4278 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4279 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4285 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4286 DebugLoc dl = Op.getDebugLoc();
4287 if (Op.getValueType() == MVT::v2f32)
4288 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4289 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4290 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4291 Op.getOperand(0))));
4293 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4294 MVT VT = MVT::v2i32;
4295 switch (Op.getValueType().getSimpleVT()) {
4302 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4303 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4306 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4307 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4308 // one of the above mentioned nodes. It has to be wrapped because otherwise
4309 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4310 // be used to form addressing mode. These wrapped nodes will be selected
4313 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4314 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4316 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4318 unsigned char OpFlag = 0;
4319 unsigned WrapperKind = X86ISD::Wrapper;
4320 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4321 if (Subtarget->isPICStyleStub())
4322 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4323 else if (Subtarget->isPICStyleGOT())
4324 OpFlag = X86II::MO_GOTOFF;
4325 else if (Subtarget->isPICStyleRIPRel() &&
4326 getTargetMachine().getCodeModel() == CodeModel::Small)
4327 WrapperKind = X86ISD::WrapperRIP;
4330 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4332 CP->getOffset(), OpFlag);
4333 DebugLoc DL = CP->getDebugLoc();
4334 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4335 // With PIC, the address is actually $g + Offset.
4337 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4338 DAG.getNode(X86ISD::GlobalBaseReg,
4339 DebugLoc::getUnknownLoc(), getPointerTy()),
4346 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4347 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4349 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4351 unsigned char OpFlag = 0;
4352 unsigned WrapperKind = X86ISD::Wrapper;
4353 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4354 if (Subtarget->isPICStyleStub())
4355 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4356 else if (Subtarget->isPICStyleGOT())
4357 OpFlag = X86II::MO_GOTOFF;
4358 else if (Subtarget->isPICStyleRIPRel())
4359 WrapperKind = X86ISD::WrapperRIP;
4362 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4364 DebugLoc DL = JT->getDebugLoc();
4365 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4367 // With PIC, the address is actually $g + Offset.
4369 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4370 DAG.getNode(X86ISD::GlobalBaseReg,
4371 DebugLoc::getUnknownLoc(), getPointerTy()),
4379 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4380 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4382 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4384 unsigned char OpFlag = 0;
4385 unsigned WrapperKind = X86ISD::Wrapper;
4386 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4387 if (Subtarget->isPICStyleStub())
4388 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4389 else if (Subtarget->isPICStyleGOT())
4390 OpFlag = X86II::MO_GOTOFF;
4391 else if (Subtarget->isPICStyleRIPRel())
4392 WrapperKind = X86ISD::WrapperRIP;
4395 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4397 DebugLoc DL = Op.getDebugLoc();
4398 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4401 // With PIC, the address is actually $g + Offset.
4402 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4403 !Subtarget->isPICStyleRIPRel()) {
4404 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4405 DAG.getNode(X86ISD::GlobalBaseReg,
4406 DebugLoc::getUnknownLoc(),
4415 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4417 SelectionDAG &DAG) const {
4418 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4419 bool ExtraLoadRequired =
4420 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4422 // Create the TargetGlobalAddress node, folding in the constant
4423 // offset if it is legal.
4425 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4426 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4429 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0);
4432 if (Subtarget->isPICStyleRIPRel() &&
4433 getTargetMachine().getCodeModel() == CodeModel::Small)
4434 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4436 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4438 // With PIC, the address is actually $g + Offset.
4439 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4440 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4441 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4445 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4446 // load the value at address GV, not the value of GV itself. This means that
4447 // the GlobalAddress must be in the base or index register of the address, not
4448 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4449 // The same applies for external symbols during PIC codegen
4450 if (ExtraLoadRequired)
4451 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4452 PseudoSourceValue::getGOT(), 0);
4454 // If there was a non-zero offset that we didn't fold, create an explicit
4457 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4458 DAG.getConstant(Offset, getPointerTy()));
4464 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4465 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4466 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4467 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4471 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4472 SDValue *InFlag, const MVT PtrVT, unsigned ReturnReg,
4473 unsigned char OperandFlags) {
4474 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4475 DebugLoc dl = GA->getDebugLoc();
4476 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4477 GA->getValueType(0),
4481 SDValue Ops[] = { Chain, TGA, *InFlag };
4482 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4484 SDValue Ops[] = { Chain, TGA };
4485 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4487 SDValue Flag = Chain.getValue(1);
4488 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4491 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4493 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4496 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4497 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4498 DAG.getNode(X86ISD::GlobalBaseReg,
4499 DebugLoc::getUnknownLoc(),
4501 InFlag = Chain.getValue(1);
4503 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4506 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4508 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4510 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4511 X86::RAX, X86II::MO_TLSGD);
4514 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4515 // "local exec" model.
4516 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4517 const MVT PtrVT, TLSModel::Model model,
4519 DebugLoc dl = GA->getDebugLoc();
4520 // Get the Thread Pointer
4521 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4522 DebugLoc::getUnknownLoc(), PtrVT,
4523 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4526 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4529 unsigned char OperandFlags = 0;
4530 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4532 unsigned WrapperKind = X86ISD::Wrapper;
4533 if (model == TLSModel::LocalExec) {
4534 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4535 } else if (is64Bit) {
4536 assert(model == TLSModel::InitialExec);
4537 OperandFlags = X86II::MO_GOTTPOFF;
4538 WrapperKind = X86ISD::WrapperRIP;
4540 assert(model == TLSModel::InitialExec);
4541 OperandFlags = X86II::MO_INDNTPOFF;
4544 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4546 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4547 GA->getOffset(), OperandFlags);
4548 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4550 if (model == TLSModel::InitialExec)
4551 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4552 PseudoSourceValue::getGOT(), 0);
4554 // The address of the thread local variable is the add of the thread
4555 // pointer with the offset of the variable.
4556 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4560 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4561 // TODO: implement the "local dynamic" model
4562 // TODO: implement the "initial exec"model for pic executables
4563 assert(Subtarget->isTargetELF() &&
4564 "TLS not implemented for non-ELF targets");
4565 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4566 const GlobalValue *GV = GA->getGlobal();
4568 // If GV is an alias then use the aliasee for determining
4569 // thread-localness.
4570 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4571 GV = GA->resolveAliasedGlobal(false);
4573 TLSModel::Model model = getTLSModel(GV,
4574 getTargetMachine().getRelocationModel());
4577 case TLSModel::GeneralDynamic:
4578 case TLSModel::LocalDynamic: // not implemented
4579 if (Subtarget->is64Bit())
4580 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4581 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4583 case TLSModel::InitialExec:
4584 case TLSModel::LocalExec:
4585 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4586 Subtarget->is64Bit());
4589 assert(0 && "Unreachable");
4594 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4595 /// take a 2 x i32 value to shift plus a shift amount.
4596 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4597 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4598 MVT VT = Op.getValueType();
4599 unsigned VTBits = VT.getSizeInBits();
4600 DebugLoc dl = Op.getDebugLoc();
4601 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4602 SDValue ShOpLo = Op.getOperand(0);
4603 SDValue ShOpHi = Op.getOperand(1);
4604 SDValue ShAmt = Op.getOperand(2);
4605 SDValue Tmp1 = isSRA ?
4606 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4607 DAG.getConstant(VTBits - 1, MVT::i8)) :
4608 DAG.getConstant(0, VT);
4611 if (Op.getOpcode() == ISD::SHL_PARTS) {
4612 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4613 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4615 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4616 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4619 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4620 DAG.getConstant(VTBits, MVT::i8));
4621 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4622 AndNode, DAG.getConstant(0, MVT::i8));
4625 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4626 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4627 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4629 if (Op.getOpcode() == ISD::SHL_PARTS) {
4630 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4631 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4633 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4634 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4637 SDValue Ops[2] = { Lo, Hi };
4638 return DAG.getMergeValues(Ops, 2, dl);
4641 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4642 MVT SrcVT = Op.getOperand(0).getValueType();
4644 if (SrcVT.isVector()) {
4645 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4651 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4652 "Unknown SINT_TO_FP to lower!");
4654 // These are really Legal; return the operand so the caller accepts it as
4656 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4658 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4659 Subtarget->is64Bit()) {
4663 DebugLoc dl = Op.getDebugLoc();
4664 unsigned Size = SrcVT.getSizeInBits()/8;
4665 MachineFunction &MF = DAG.getMachineFunction();
4666 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4667 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4668 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4670 PseudoSourceValue::getFixedStack(SSFI), 0);
4671 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4674 SDValue X86TargetLowering::BuildFILD(SDValue Op, MVT SrcVT, SDValue Chain,
4676 SelectionDAG &DAG) {
4678 DebugLoc dl = Op.getDebugLoc();
4680 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4682 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4684 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4685 SmallVector<SDValue, 8> Ops;
4686 Ops.push_back(Chain);
4687 Ops.push_back(StackSlot);
4688 Ops.push_back(DAG.getValueType(SrcVT));
4689 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4690 Tys, &Ops[0], Ops.size());
4693 Chain = Result.getValue(1);
4694 SDValue InFlag = Result.getValue(2);
4696 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4697 // shouldn't be necessary except that RFP cannot be live across
4698 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4699 MachineFunction &MF = DAG.getMachineFunction();
4700 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4701 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4702 Tys = DAG.getVTList(MVT::Other);
4703 SmallVector<SDValue, 8> Ops;
4704 Ops.push_back(Chain);
4705 Ops.push_back(Result);
4706 Ops.push_back(StackSlot);
4707 Ops.push_back(DAG.getValueType(Op.getValueType()));
4708 Ops.push_back(InFlag);
4709 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4710 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4711 PseudoSourceValue::getFixedStack(SSFI), 0);
4717 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4718 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4719 // This algorithm is not obvious. Here it is in C code, more or less:
4721 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4722 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4723 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4725 // Copy ints to xmm registers.
4726 __m128i xh = _mm_cvtsi32_si128( hi );
4727 __m128i xl = _mm_cvtsi32_si128( lo );
4729 // Combine into low half of a single xmm register.
4730 __m128i x = _mm_unpacklo_epi32( xh, xl );
4734 // Merge in appropriate exponents to give the integer bits the right
4736 x = _mm_unpacklo_epi32( x, exp );
4738 // Subtract away the biases to deal with the IEEE-754 double precision
4740 d = _mm_sub_pd( (__m128d) x, bias );
4742 // All conversions up to here are exact. The correctly rounded result is
4743 // calculated using the current rounding mode using the following
4745 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4746 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4747 // store doesn't really need to be here (except
4748 // maybe to zero the other double)
4753 DebugLoc dl = Op.getDebugLoc();
4755 // Build some magic constants.
4756 std::vector<Constant*> CV0;
4757 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
4758 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
4759 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4760 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4761 Constant *C0 = ConstantVector::get(CV0);
4762 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4764 std::vector<Constant*> CV1;
4765 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
4766 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
4767 Constant *C1 = ConstantVector::get(CV1);
4768 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4770 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4771 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4773 DAG.getIntPtrConstant(1)));
4774 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4775 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4777 DAG.getIntPtrConstant(0)));
4778 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4779 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4780 PseudoSourceValue::getConstantPool(), 0,
4782 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4783 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4784 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4785 PseudoSourceValue::getConstantPool(), 0,
4787 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4789 // Add the halves; easiest way is to swap them into another reg first.
4790 int ShufMask[2] = { 1, -1 };
4791 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4792 DAG.getUNDEF(MVT::v2f64), ShufMask);
4793 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4794 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4795 DAG.getIntPtrConstant(0));
4798 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4799 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4800 DebugLoc dl = Op.getDebugLoc();
4801 // FP constant to bias correct the final result.
4802 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4805 // Load the 32-bit value into an XMM register.
4806 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4807 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4809 DAG.getIntPtrConstant(0)));
4811 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4812 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4813 DAG.getIntPtrConstant(0));
4815 // Or the load with the bias.
4816 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4817 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4818 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4820 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4821 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4822 MVT::v2f64, Bias)));
4823 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4824 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
4825 DAG.getIntPtrConstant(0));
4827 // Subtract the bias.
4828 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
4830 // Handle final rounding.
4831 MVT DestVT = Op.getValueType();
4833 if (DestVT.bitsLT(MVT::f64)) {
4834 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
4835 DAG.getIntPtrConstant(0));
4836 } else if (DestVT.bitsGT(MVT::f64)) {
4837 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
4840 // Handle final rounding.
4844 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4845 SDValue N0 = Op.getOperand(0);
4846 DebugLoc dl = Op.getDebugLoc();
4848 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
4849 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
4850 // the optimization here.
4851 if (DAG.SignBitIsZero(N0))
4852 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
4854 MVT SrcVT = N0.getValueType();
4855 if (SrcVT == MVT::i64) {
4856 // We only handle SSE2 f64 target here; caller can expand the rest.
4857 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
4860 return LowerUINT_TO_FP_i64(Op, DAG);
4861 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
4862 return LowerUINT_TO_FP_i32(Op, DAG);
4865 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
4867 // Make a 64-bit buffer, and use it to build an FILD.
4868 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
4869 SDValue WordOff = DAG.getConstant(4, getPointerTy());
4870 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
4871 getPointerTy(), StackSlot, WordOff);
4872 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4873 StackSlot, NULL, 0);
4874 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
4875 OffsetSlot, NULL, 0);
4876 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
4879 std::pair<SDValue,SDValue> X86TargetLowering::
4880 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
4881 DebugLoc dl = Op.getDebugLoc();
4883 MVT DstTy = Op.getValueType();
4886 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
4890 assert(DstTy.getSimpleVT() <= MVT::i64 &&
4891 DstTy.getSimpleVT() >= MVT::i16 &&
4892 "Unknown FP_TO_SINT to lower!");
4894 // These are really Legal.
4895 if (DstTy == MVT::i32 &&
4896 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4897 return std::make_pair(SDValue(), SDValue());
4898 if (Subtarget->is64Bit() &&
4899 DstTy == MVT::i64 &&
4900 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4901 return std::make_pair(SDValue(), SDValue());
4903 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4905 MachineFunction &MF = DAG.getMachineFunction();
4906 unsigned MemSize = DstTy.getSizeInBits()/8;
4907 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4908 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4911 switch (DstTy.getSimpleVT()) {
4912 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4913 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4914 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4915 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4918 SDValue Chain = DAG.getEntryNode();
4919 SDValue Value = Op.getOperand(0);
4920 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4921 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4922 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
4923 PseudoSourceValue::getFixedStack(SSFI), 0);
4924 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4926 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4928 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
4929 Chain = Value.getValue(1);
4930 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4931 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4934 // Build the FP_TO_INT*_IN_MEM
4935 SDValue Ops[] = { Chain, Value, StackSlot };
4936 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
4938 return std::make_pair(FIST, StackSlot);
4941 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4942 if (Op.getValueType().isVector()) {
4943 if (Op.getValueType() == MVT::v2i32 &&
4944 Op.getOperand(0).getValueType() == MVT::v2f64) {
4950 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
4951 SDValue FIST = Vals.first, StackSlot = Vals.second;
4952 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
4953 if (FIST.getNode() == 0) return Op;
4956 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
4957 FIST, StackSlot, NULL, 0);
4960 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
4961 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
4962 SDValue FIST = Vals.first, StackSlot = Vals.second;
4963 assert(FIST.getNode() && "Unexpected failure");
4966 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
4967 FIST, StackSlot, NULL, 0);
4970 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4971 DebugLoc dl = Op.getDebugLoc();
4972 MVT VT = Op.getValueType();
4975 EltVT = VT.getVectorElementType();
4976 std::vector<Constant*> CV;
4977 if (EltVT == MVT::f64) {
4978 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4982 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4988 Constant *C = ConstantVector::get(CV);
4989 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
4990 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
4991 PseudoSourceValue::getConstantPool(), 0,
4993 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
4996 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4997 DebugLoc dl = Op.getDebugLoc();
4998 MVT VT = Op.getValueType();
5000 unsigned EltNum = 1;
5001 if (VT.isVector()) {
5002 EltVT = VT.getVectorElementType();
5003 EltNum = VT.getVectorNumElements();
5005 std::vector<Constant*> CV;
5006 if (EltVT == MVT::f64) {
5007 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
5011 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
5017 Constant *C = ConstantVector::get(CV);
5018 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5019 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5020 PseudoSourceValue::getConstantPool(), 0,
5022 if (VT.isVector()) {
5023 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5024 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5025 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5027 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5029 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5033 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5034 SDValue Op0 = Op.getOperand(0);
5035 SDValue Op1 = Op.getOperand(1);
5036 DebugLoc dl = Op.getDebugLoc();
5037 MVT VT = Op.getValueType();
5038 MVT SrcVT = Op1.getValueType();
5040 // If second operand is smaller, extend it first.
5041 if (SrcVT.bitsLT(VT)) {
5042 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5045 // And if it is bigger, shrink it first.
5046 if (SrcVT.bitsGT(VT)) {
5047 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5051 // At this point the operands and the result should have the same
5052 // type, and that won't be f80 since that is not custom lowered.
5054 // First get the sign bit of second operand.
5055 std::vector<Constant*> CV;
5056 if (SrcVT == MVT::f64) {
5057 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
5058 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5060 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
5061 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5062 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5063 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5065 Constant *C = ConstantVector::get(CV);
5066 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5067 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5068 PseudoSourceValue::getConstantPool(), 0,
5070 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5072 // Shift sign bit right or left if the two operands have different types.
5073 if (SrcVT.bitsGT(VT)) {
5074 // Op0 is MVT::f32, Op1 is MVT::f64.
5075 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5076 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5077 DAG.getConstant(32, MVT::i32));
5078 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5079 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5080 DAG.getIntPtrConstant(0));
5083 // Clear first operand sign bit.
5085 if (VT == MVT::f64) {
5086 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5087 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5089 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5090 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5091 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5092 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5094 C = ConstantVector::get(CV);
5095 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5096 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5097 PseudoSourceValue::getConstantPool(), 0,
5099 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5101 // Or the value with the sign bit.
5102 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5105 /// Emit nodes that will be selected as "test Op0,Op0", or something
5107 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5108 SelectionDAG &DAG) {
5109 DebugLoc dl = Op.getDebugLoc();
5111 // CF and OF aren't always set the way we want. Determine which
5112 // of these we need.
5113 bool NeedCF = false;
5114 bool NeedOF = false;
5116 case X86::COND_A: case X86::COND_AE:
5117 case X86::COND_B: case X86::COND_BE:
5120 case X86::COND_G: case X86::COND_GE:
5121 case X86::COND_L: case X86::COND_LE:
5122 case X86::COND_O: case X86::COND_NO:
5128 // See if we can use the EFLAGS value from the operand instead of
5129 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5130 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5131 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5132 unsigned Opcode = 0;
5133 unsigned NumOperands = 0;
5134 switch (Op.getNode()->getOpcode()) {
5136 // Due to an isel shortcoming, be conservative if this add is likely to
5137 // be selected as part of a load-modify-store instruction. When the root
5138 // node in a match is a store, isel doesn't know how to remap non-chain
5139 // non-flag uses of other nodes in the match, such as the ADD in this
5140 // case. This leads to the ADD being left around and reselected, with
5141 // the result being two adds in the output.
5142 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5143 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5144 if (UI->getOpcode() == ISD::STORE)
5146 if (ConstantSDNode *C =
5147 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5148 // An add of one will be selected as an INC.
5149 if (C->getAPIntValue() == 1) {
5150 Opcode = X86ISD::INC;
5154 // An add of negative one (subtract of one) will be selected as a DEC.
5155 if (C->getAPIntValue().isAllOnesValue()) {
5156 Opcode = X86ISD::DEC;
5161 // Otherwise use a regular EFLAGS-setting add.
5162 Opcode = X86ISD::ADD;
5166 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5167 // likely to be selected as part of a load-modify-store instruction.
5168 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5169 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5170 if (UI->getOpcode() == ISD::STORE)
5172 // Otherwise use a regular EFLAGS-setting sub.
5173 Opcode = X86ISD::SUB;
5180 return SDValue(Op.getNode(), 1);
5186 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5187 SmallVector<SDValue, 4> Ops;
5188 for (unsigned i = 0; i != NumOperands; ++i)
5189 Ops.push_back(Op.getOperand(i));
5190 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5191 DAG.ReplaceAllUsesWith(Op, New);
5192 return SDValue(New.getNode(), 1);
5196 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5197 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5198 DAG.getConstant(0, Op.getValueType()));
5201 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5203 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5204 SelectionDAG &DAG) {
5205 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5206 if (C->getAPIntValue() == 0)
5207 return EmitTest(Op0, X86CC, DAG);
5209 DebugLoc dl = Op0.getDebugLoc();
5210 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5213 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5214 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5215 SDValue Op0 = Op.getOperand(0);
5216 SDValue Op1 = Op.getOperand(1);
5217 DebugLoc dl = Op.getDebugLoc();
5218 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5220 // Lower (X & (1 << N)) == 0 to BT(X, N).
5221 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5222 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5223 if (Op0.getOpcode() == ISD::AND &&
5225 Op1.getOpcode() == ISD::Constant &&
5226 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5227 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5229 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5230 if (ConstantSDNode *Op010C =
5231 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5232 if (Op010C->getZExtValue() == 1) {
5233 LHS = Op0.getOperand(0);
5234 RHS = Op0.getOperand(1).getOperand(1);
5236 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5237 if (ConstantSDNode *Op000C =
5238 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5239 if (Op000C->getZExtValue() == 1) {
5240 LHS = Op0.getOperand(1);
5241 RHS = Op0.getOperand(0).getOperand(1);
5243 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5244 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5245 SDValue AndLHS = Op0.getOperand(0);
5246 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5247 LHS = AndLHS.getOperand(0);
5248 RHS = AndLHS.getOperand(1);
5252 if (LHS.getNode()) {
5253 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5254 // instruction. Since the shift amount is in-range-or-undefined, we know
5255 // that doing a bittest on the i16 value is ok. We extend to i32 because
5256 // the encoding for the i16 version is larger than the i32 version.
5257 if (LHS.getValueType() == MVT::i8)
5258 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5260 // If the operand types disagree, extend the shift amount to match. Since
5261 // BT ignores high bits (like shifts) we can use anyextend.
5262 if (LHS.getValueType() != RHS.getValueType())
5263 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5265 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5266 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5267 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5268 DAG.getConstant(Cond, MVT::i8), BT);
5272 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5273 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5275 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5276 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5277 DAG.getConstant(X86CC, MVT::i8), Cond);
5280 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5282 SDValue Op0 = Op.getOperand(0);
5283 SDValue Op1 = Op.getOperand(1);
5284 SDValue CC = Op.getOperand(2);
5285 MVT VT = Op.getValueType();
5286 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5287 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5288 DebugLoc dl = Op.getDebugLoc();
5292 MVT VT0 = Op0.getValueType();
5293 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5294 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5297 switch (SetCCOpcode) {
5300 case ISD::SETEQ: SSECC = 0; break;
5302 case ISD::SETGT: Swap = true; // Fallthrough
5304 case ISD::SETOLT: SSECC = 1; break;
5306 case ISD::SETGE: Swap = true; // Fallthrough
5308 case ISD::SETOLE: SSECC = 2; break;
5309 case ISD::SETUO: SSECC = 3; break;
5311 case ISD::SETNE: SSECC = 4; break;
5312 case ISD::SETULE: Swap = true;
5313 case ISD::SETUGE: SSECC = 5; break;
5314 case ISD::SETULT: Swap = true;
5315 case ISD::SETUGT: SSECC = 6; break;
5316 case ISD::SETO: SSECC = 7; break;
5319 std::swap(Op0, Op1);
5321 // In the two special cases we can't handle, emit two comparisons.
5323 if (SetCCOpcode == ISD::SETUEQ) {
5325 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5326 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5327 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5329 else if (SetCCOpcode == ISD::SETONE) {
5331 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5332 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5333 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5335 assert(0 && "Illegal FP comparison");
5337 // Handle all other FP comparisons here.
5338 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5341 // We are handling one of the integer comparisons here. Since SSE only has
5342 // GT and EQ comparisons for integer, swapping operands and multiple
5343 // operations may be required for some comparisons.
5344 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5345 bool Swap = false, Invert = false, FlipSigns = false;
5347 switch (VT.getSimpleVT()) {
5349 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5350 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5351 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5352 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5355 switch (SetCCOpcode) {
5357 case ISD::SETNE: Invert = true;
5358 case ISD::SETEQ: Opc = EQOpc; break;
5359 case ISD::SETLT: Swap = true;
5360 case ISD::SETGT: Opc = GTOpc; break;
5361 case ISD::SETGE: Swap = true;
5362 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5363 case ISD::SETULT: Swap = true;
5364 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5365 case ISD::SETUGE: Swap = true;
5366 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5369 std::swap(Op0, Op1);
5371 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5372 // bits of the inputs before performing those operations.
5374 MVT EltVT = VT.getVectorElementType();
5375 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5377 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5378 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5380 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5381 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5384 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5386 // If the logical-not of the result is required, perform that now.
5388 Result = DAG.getNOT(dl, Result, VT);
5393 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5394 static bool isX86LogicalCmp(SDValue Op) {
5395 unsigned Opc = Op.getNode()->getOpcode();
5396 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5398 if (Op.getResNo() == 1 &&
5399 (Opc == X86ISD::ADD ||
5400 Opc == X86ISD::SUB ||
5401 Opc == X86ISD::SMUL ||
5402 Opc == X86ISD::UMUL ||
5403 Opc == X86ISD::INC ||
5404 Opc == X86ISD::DEC))
5410 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5411 bool addTest = true;
5412 SDValue Cond = Op.getOperand(0);
5413 DebugLoc dl = Op.getDebugLoc();
5416 if (Cond.getOpcode() == ISD::SETCC)
5417 Cond = LowerSETCC(Cond, DAG);
5419 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5420 // setting operand in place of the X86ISD::SETCC.
5421 if (Cond.getOpcode() == X86ISD::SETCC) {
5422 CC = Cond.getOperand(0);
5424 SDValue Cmp = Cond.getOperand(1);
5425 unsigned Opc = Cmp.getOpcode();
5426 MVT VT = Op.getValueType();
5428 bool IllegalFPCMov = false;
5429 if (VT.isFloatingPoint() && !VT.isVector() &&
5430 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5431 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5433 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5434 Opc == X86ISD::BT) { // FIXME
5441 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5442 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5445 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5446 SmallVector<SDValue, 4> Ops;
5447 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5448 // condition is true.
5449 Ops.push_back(Op.getOperand(2));
5450 Ops.push_back(Op.getOperand(1));
5452 Ops.push_back(Cond);
5453 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5456 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5457 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5458 // from the AND / OR.
5459 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5460 Opc = Op.getOpcode();
5461 if (Opc != ISD::OR && Opc != ISD::AND)
5463 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5464 Op.getOperand(0).hasOneUse() &&
5465 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5466 Op.getOperand(1).hasOneUse());
5469 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5470 // 1 and that the SETCC node has a single use.
5471 static bool isXor1OfSetCC(SDValue Op) {
5472 if (Op.getOpcode() != ISD::XOR)
5474 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5475 if (N1C && N1C->getAPIntValue() == 1) {
5476 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5477 Op.getOperand(0).hasOneUse();
5482 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5483 bool addTest = true;
5484 SDValue Chain = Op.getOperand(0);
5485 SDValue Cond = Op.getOperand(1);
5486 SDValue Dest = Op.getOperand(2);
5487 DebugLoc dl = Op.getDebugLoc();
5490 if (Cond.getOpcode() == ISD::SETCC)
5491 Cond = LowerSETCC(Cond, DAG);
5493 // FIXME: LowerXALUO doesn't handle these!!
5494 else if (Cond.getOpcode() == X86ISD::ADD ||
5495 Cond.getOpcode() == X86ISD::SUB ||
5496 Cond.getOpcode() == X86ISD::SMUL ||
5497 Cond.getOpcode() == X86ISD::UMUL)
5498 Cond = LowerXALUO(Cond, DAG);
5501 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5502 // setting operand in place of the X86ISD::SETCC.
5503 if (Cond.getOpcode() == X86ISD::SETCC) {
5504 CC = Cond.getOperand(0);
5506 SDValue Cmp = Cond.getOperand(1);
5507 unsigned Opc = Cmp.getOpcode();
5508 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5509 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5513 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5517 // These can only come from an arithmetic instruction with overflow,
5518 // e.g. SADDO, UADDO.
5519 Cond = Cond.getNode()->getOperand(1);
5526 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5527 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5528 if (CondOpc == ISD::OR) {
5529 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5530 // two branches instead of an explicit OR instruction with a
5532 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5533 isX86LogicalCmp(Cmp)) {
5534 CC = Cond.getOperand(0).getOperand(0);
5535 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5536 Chain, Dest, CC, Cmp);
5537 CC = Cond.getOperand(1).getOperand(0);
5541 } else { // ISD::AND
5542 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5543 // two branches instead of an explicit AND instruction with a
5544 // separate test. However, we only do this if this block doesn't
5545 // have a fall-through edge, because this requires an explicit
5546 // jmp when the condition is false.
5547 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5548 isX86LogicalCmp(Cmp) &&
5549 Op.getNode()->hasOneUse()) {
5550 X86::CondCode CCode =
5551 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5552 CCode = X86::GetOppositeBranchCondition(CCode);
5553 CC = DAG.getConstant(CCode, MVT::i8);
5554 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5555 // Look for an unconditional branch following this conditional branch.
5556 // We need this because we need to reverse the successors in order
5557 // to implement FCMP_OEQ.
5558 if (User.getOpcode() == ISD::BR) {
5559 SDValue FalseBB = User.getOperand(1);
5561 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5562 assert(NewBR == User);
5565 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5566 Chain, Dest, CC, Cmp);
5567 X86::CondCode CCode =
5568 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5569 CCode = X86::GetOppositeBranchCondition(CCode);
5570 CC = DAG.getConstant(CCode, MVT::i8);
5576 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5577 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5578 // It should be transformed during dag combiner except when the condition
5579 // is set by a arithmetics with overflow node.
5580 X86::CondCode CCode =
5581 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5582 CCode = X86::GetOppositeBranchCondition(CCode);
5583 CC = DAG.getConstant(CCode, MVT::i8);
5584 Cond = Cond.getOperand(0).getOperand(1);
5590 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5591 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5593 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5594 Chain, Dest, CC, Cond);
5598 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5599 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5600 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5601 // that the guard pages used by the OS virtual memory manager are allocated in
5602 // correct sequence.
5604 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5605 SelectionDAG &DAG) {
5606 assert(Subtarget->isTargetCygMing() &&
5607 "This should be used only on Cygwin/Mingw targets");
5608 DebugLoc dl = Op.getDebugLoc();
5611 SDValue Chain = Op.getOperand(0);
5612 SDValue Size = Op.getOperand(1);
5613 // FIXME: Ensure alignment here
5617 MVT IntPtr = getPointerTy();
5618 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5620 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5622 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5623 Flag = Chain.getValue(1);
5625 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5626 SDValue Ops[] = { Chain,
5627 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5628 DAG.getRegister(X86::EAX, IntPtr),
5629 DAG.getRegister(X86StackPtr, SPTy),
5631 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5632 Flag = Chain.getValue(1);
5634 Chain = DAG.getCALLSEQ_END(Chain,
5635 DAG.getIntPtrConstant(0, true),
5636 DAG.getIntPtrConstant(0, true),
5639 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5641 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5642 return DAG.getMergeValues(Ops1, 2, dl);
5646 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5648 SDValue Dst, SDValue Src,
5649 SDValue Size, unsigned Align,
5651 uint64_t DstSVOff) {
5652 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5654 // If not DWORD aligned or size is more than the threshold, call the library.
5655 // The libc version is likely to be faster for these cases. It can use the
5656 // address value and run time information about the CPU.
5657 if ((Align & 3) != 0 ||
5659 ConstantSize->getZExtValue() >
5660 getSubtarget()->getMaxInlineSizeThreshold()) {
5661 SDValue InFlag(0, 0);
5663 // Check to see if there is a specialized entry-point for memory zeroing.
5664 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5666 if (const char *bzeroEntry = V &&
5667 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5668 MVT IntPtr = getPointerTy();
5669 const Type *IntPtrTy = TD->getIntPtrType();
5670 TargetLowering::ArgListTy Args;
5671 TargetLowering::ArgListEntry Entry;
5673 Entry.Ty = IntPtrTy;
5674 Args.push_back(Entry);
5676 Args.push_back(Entry);
5677 std::pair<SDValue,SDValue> CallResult =
5678 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5679 CallingConv::C, false,
5680 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5681 return CallResult.second;
5684 // Otherwise have the target-independent code call memset.
5688 uint64_t SizeVal = ConstantSize->getZExtValue();
5689 SDValue InFlag(0, 0);
5692 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5693 unsigned BytesLeft = 0;
5694 bool TwoRepStos = false;
5697 uint64_t Val = ValC->getZExtValue() & 255;
5699 // If the value is a constant, then we can potentially use larger sets.
5700 switch (Align & 3) {
5701 case 2: // WORD aligned
5704 Val = (Val << 8) | Val;
5706 case 0: // DWORD aligned
5709 Val = (Val << 8) | Val;
5710 Val = (Val << 16) | Val;
5711 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5714 Val = (Val << 32) | Val;
5717 default: // Byte aligned
5720 Count = DAG.getIntPtrConstant(SizeVal);
5724 if (AVT.bitsGT(MVT::i8)) {
5725 unsigned UBytes = AVT.getSizeInBits() / 8;
5726 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5727 BytesLeft = SizeVal % UBytes;
5730 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5732 InFlag = Chain.getValue(1);
5735 Count = DAG.getIntPtrConstant(SizeVal);
5736 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5737 InFlag = Chain.getValue(1);
5740 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5743 InFlag = Chain.getValue(1);
5744 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5747 InFlag = Chain.getValue(1);
5749 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5750 SmallVector<SDValue, 8> Ops;
5751 Ops.push_back(Chain);
5752 Ops.push_back(DAG.getValueType(AVT));
5753 Ops.push_back(InFlag);
5754 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5757 InFlag = Chain.getValue(1);
5759 MVT CVT = Count.getValueType();
5760 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5761 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5762 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5765 InFlag = Chain.getValue(1);
5766 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5768 Ops.push_back(Chain);
5769 Ops.push_back(DAG.getValueType(MVT::i8));
5770 Ops.push_back(InFlag);
5771 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5772 } else if (BytesLeft) {
5773 // Handle the last 1 - 7 bytes.
5774 unsigned Offset = SizeVal - BytesLeft;
5775 MVT AddrVT = Dst.getValueType();
5776 MVT SizeVT = Size.getValueType();
5778 Chain = DAG.getMemset(Chain, dl,
5779 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5780 DAG.getConstant(Offset, AddrVT)),
5782 DAG.getConstant(BytesLeft, SizeVT),
5783 Align, DstSV, DstSVOff + Offset);
5786 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5791 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5792 SDValue Chain, SDValue Dst, SDValue Src,
5793 SDValue Size, unsigned Align,
5795 const Value *DstSV, uint64_t DstSVOff,
5796 const Value *SrcSV, uint64_t SrcSVOff) {
5797 // This requires the copy size to be a constant, preferrably
5798 // within a subtarget-specific limit.
5799 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5802 uint64_t SizeVal = ConstantSize->getZExtValue();
5803 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5806 /// If not DWORD aligned, call the library.
5807 if ((Align & 3) != 0)
5812 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5815 unsigned UBytes = AVT.getSizeInBits() / 8;
5816 unsigned CountVal = SizeVal / UBytes;
5817 SDValue Count = DAG.getIntPtrConstant(CountVal);
5818 unsigned BytesLeft = SizeVal % UBytes;
5820 SDValue InFlag(0, 0);
5821 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5824 InFlag = Chain.getValue(1);
5825 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5828 InFlag = Chain.getValue(1);
5829 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
5832 InFlag = Chain.getValue(1);
5834 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5835 SmallVector<SDValue, 8> Ops;
5836 Ops.push_back(Chain);
5837 Ops.push_back(DAG.getValueType(AVT));
5838 Ops.push_back(InFlag);
5839 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
5841 SmallVector<SDValue, 4> Results;
5842 Results.push_back(RepMovs);
5844 // Handle the last 1 - 7 bytes.
5845 unsigned Offset = SizeVal - BytesLeft;
5846 MVT DstVT = Dst.getValueType();
5847 MVT SrcVT = Src.getValueType();
5848 MVT SizeVT = Size.getValueType();
5849 Results.push_back(DAG.getMemcpy(Chain, dl,
5850 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
5851 DAG.getConstant(Offset, DstVT)),
5852 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
5853 DAG.getConstant(Offset, SrcVT)),
5854 DAG.getConstant(BytesLeft, SizeVT),
5855 Align, AlwaysInline,
5856 DstSV, DstSVOff + Offset,
5857 SrcSV, SrcSVOff + Offset));
5860 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5861 &Results[0], Results.size());
5864 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5865 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5866 DebugLoc dl = Op.getDebugLoc();
5868 if (!Subtarget->is64Bit()) {
5869 // vastart just stores the address of the VarArgsFrameIndex slot into the
5870 // memory location argument.
5871 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5872 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
5876 // gp_offset (0 - 6 * 8)
5877 // fp_offset (48 - 48 + 8 * 16)
5878 // overflow_arg_area (point to parameters coming in memory).
5880 SmallVector<SDValue, 8> MemOps;
5881 SDValue FIN = Op.getOperand(1);
5883 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
5884 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5886 MemOps.push_back(Store);
5889 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5890 FIN, DAG.getIntPtrConstant(4));
5891 Store = DAG.getStore(Op.getOperand(0), dl,
5892 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5894 MemOps.push_back(Store);
5896 // Store ptr to overflow_arg_area
5897 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5898 FIN, DAG.getIntPtrConstant(4));
5899 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5900 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
5901 MemOps.push_back(Store);
5903 // Store ptr to reg_save_area.
5904 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5905 FIN, DAG.getIntPtrConstant(8));
5906 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5907 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
5908 MemOps.push_back(Store);
5909 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5910 &MemOps[0], MemOps.size());
5913 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5914 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5915 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5916 SDValue Chain = Op.getOperand(0);
5917 SDValue SrcPtr = Op.getOperand(1);
5918 SDValue SrcSV = Op.getOperand(2);
5920 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5925 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5926 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5927 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5928 SDValue Chain = Op.getOperand(0);
5929 SDValue DstPtr = Op.getOperand(1);
5930 SDValue SrcPtr = Op.getOperand(2);
5931 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5932 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5933 DebugLoc dl = Op.getDebugLoc();
5935 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
5936 DAG.getIntPtrConstant(24), 8, false,
5937 DstSV, 0, SrcSV, 0);
5941 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5942 DebugLoc dl = Op.getDebugLoc();
5943 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5945 default: return SDValue(); // Don't custom lower most intrinsics.
5946 // Comparison intrinsics.
5947 case Intrinsic::x86_sse_comieq_ss:
5948 case Intrinsic::x86_sse_comilt_ss:
5949 case Intrinsic::x86_sse_comile_ss:
5950 case Intrinsic::x86_sse_comigt_ss:
5951 case Intrinsic::x86_sse_comige_ss:
5952 case Intrinsic::x86_sse_comineq_ss:
5953 case Intrinsic::x86_sse_ucomieq_ss:
5954 case Intrinsic::x86_sse_ucomilt_ss:
5955 case Intrinsic::x86_sse_ucomile_ss:
5956 case Intrinsic::x86_sse_ucomigt_ss:
5957 case Intrinsic::x86_sse_ucomige_ss:
5958 case Intrinsic::x86_sse_ucomineq_ss:
5959 case Intrinsic::x86_sse2_comieq_sd:
5960 case Intrinsic::x86_sse2_comilt_sd:
5961 case Intrinsic::x86_sse2_comile_sd:
5962 case Intrinsic::x86_sse2_comigt_sd:
5963 case Intrinsic::x86_sse2_comige_sd:
5964 case Intrinsic::x86_sse2_comineq_sd:
5965 case Intrinsic::x86_sse2_ucomieq_sd:
5966 case Intrinsic::x86_sse2_ucomilt_sd:
5967 case Intrinsic::x86_sse2_ucomile_sd:
5968 case Intrinsic::x86_sse2_ucomigt_sd:
5969 case Intrinsic::x86_sse2_ucomige_sd:
5970 case Intrinsic::x86_sse2_ucomineq_sd: {
5972 ISD::CondCode CC = ISD::SETCC_INVALID;
5975 case Intrinsic::x86_sse_comieq_ss:
5976 case Intrinsic::x86_sse2_comieq_sd:
5980 case Intrinsic::x86_sse_comilt_ss:
5981 case Intrinsic::x86_sse2_comilt_sd:
5985 case Intrinsic::x86_sse_comile_ss:
5986 case Intrinsic::x86_sse2_comile_sd:
5990 case Intrinsic::x86_sse_comigt_ss:
5991 case Intrinsic::x86_sse2_comigt_sd:
5995 case Intrinsic::x86_sse_comige_ss:
5996 case Intrinsic::x86_sse2_comige_sd:
6000 case Intrinsic::x86_sse_comineq_ss:
6001 case Intrinsic::x86_sse2_comineq_sd:
6005 case Intrinsic::x86_sse_ucomieq_ss:
6006 case Intrinsic::x86_sse2_ucomieq_sd:
6007 Opc = X86ISD::UCOMI;
6010 case Intrinsic::x86_sse_ucomilt_ss:
6011 case Intrinsic::x86_sse2_ucomilt_sd:
6012 Opc = X86ISD::UCOMI;
6015 case Intrinsic::x86_sse_ucomile_ss:
6016 case Intrinsic::x86_sse2_ucomile_sd:
6017 Opc = X86ISD::UCOMI;
6020 case Intrinsic::x86_sse_ucomigt_ss:
6021 case Intrinsic::x86_sse2_ucomigt_sd:
6022 Opc = X86ISD::UCOMI;
6025 case Intrinsic::x86_sse_ucomige_ss:
6026 case Intrinsic::x86_sse2_ucomige_sd:
6027 Opc = X86ISD::UCOMI;
6030 case Intrinsic::x86_sse_ucomineq_ss:
6031 case Intrinsic::x86_sse2_ucomineq_sd:
6032 Opc = X86ISD::UCOMI;
6037 SDValue LHS = Op.getOperand(1);
6038 SDValue RHS = Op.getOperand(2);
6039 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6040 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6041 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6042 DAG.getConstant(X86CC, MVT::i8), Cond);
6043 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6046 // Fix vector shift instructions where the last operand is a non-immediate
6048 case Intrinsic::x86_sse2_pslli_w:
6049 case Intrinsic::x86_sse2_pslli_d:
6050 case Intrinsic::x86_sse2_pslli_q:
6051 case Intrinsic::x86_sse2_psrli_w:
6052 case Intrinsic::x86_sse2_psrli_d:
6053 case Intrinsic::x86_sse2_psrli_q:
6054 case Intrinsic::x86_sse2_psrai_w:
6055 case Intrinsic::x86_sse2_psrai_d:
6056 case Intrinsic::x86_mmx_pslli_w:
6057 case Intrinsic::x86_mmx_pslli_d:
6058 case Intrinsic::x86_mmx_pslli_q:
6059 case Intrinsic::x86_mmx_psrli_w:
6060 case Intrinsic::x86_mmx_psrli_d:
6061 case Intrinsic::x86_mmx_psrli_q:
6062 case Intrinsic::x86_mmx_psrai_w:
6063 case Intrinsic::x86_mmx_psrai_d: {
6064 SDValue ShAmt = Op.getOperand(2);
6065 if (isa<ConstantSDNode>(ShAmt))
6068 unsigned NewIntNo = 0;
6069 MVT ShAmtVT = MVT::v4i32;
6071 case Intrinsic::x86_sse2_pslli_w:
6072 NewIntNo = Intrinsic::x86_sse2_psll_w;
6074 case Intrinsic::x86_sse2_pslli_d:
6075 NewIntNo = Intrinsic::x86_sse2_psll_d;
6077 case Intrinsic::x86_sse2_pslli_q:
6078 NewIntNo = Intrinsic::x86_sse2_psll_q;
6080 case Intrinsic::x86_sse2_psrli_w:
6081 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6083 case Intrinsic::x86_sse2_psrli_d:
6084 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6086 case Intrinsic::x86_sse2_psrli_q:
6087 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6089 case Intrinsic::x86_sse2_psrai_w:
6090 NewIntNo = Intrinsic::x86_sse2_psra_w;
6092 case Intrinsic::x86_sse2_psrai_d:
6093 NewIntNo = Intrinsic::x86_sse2_psra_d;
6096 ShAmtVT = MVT::v2i32;
6098 case Intrinsic::x86_mmx_pslli_w:
6099 NewIntNo = Intrinsic::x86_mmx_psll_w;
6101 case Intrinsic::x86_mmx_pslli_d:
6102 NewIntNo = Intrinsic::x86_mmx_psll_d;
6104 case Intrinsic::x86_mmx_pslli_q:
6105 NewIntNo = Intrinsic::x86_mmx_psll_q;
6107 case Intrinsic::x86_mmx_psrli_w:
6108 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6110 case Intrinsic::x86_mmx_psrli_d:
6111 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6113 case Intrinsic::x86_mmx_psrli_q:
6114 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6116 case Intrinsic::x86_mmx_psrai_w:
6117 NewIntNo = Intrinsic::x86_mmx_psra_w;
6119 case Intrinsic::x86_mmx_psrai_d:
6120 NewIntNo = Intrinsic::x86_mmx_psra_d;
6122 default: abort(); // Can't reach here.
6127 MVT VT = Op.getValueType();
6128 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6129 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6130 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6131 DAG.getConstant(NewIntNo, MVT::i32),
6132 Op.getOperand(1), ShAmt);
6137 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6138 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6139 DebugLoc dl = Op.getDebugLoc();
6142 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6144 DAG.getConstant(TD->getPointerSize(),
6145 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6146 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6147 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6152 // Just load the return address.
6153 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6154 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6155 RetAddrFI, NULL, 0);
6158 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6159 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6160 MFI->setFrameAddressIsTaken(true);
6161 MVT VT = Op.getValueType();
6162 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6163 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6164 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6165 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6167 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6171 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6172 SelectionDAG &DAG) {
6173 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6176 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6178 MachineFunction &MF = DAG.getMachineFunction();
6179 SDValue Chain = Op.getOperand(0);
6180 SDValue Offset = Op.getOperand(1);
6181 SDValue Handler = Op.getOperand(2);
6182 DebugLoc dl = Op.getDebugLoc();
6184 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6186 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6188 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6189 DAG.getIntPtrConstant(-TD->getPointerSize()));
6190 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6191 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6192 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6193 MF.getRegInfo().addLiveOut(StoreAddrReg);
6195 return DAG.getNode(X86ISD::EH_RETURN, dl,
6197 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6200 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6201 SelectionDAG &DAG) {
6202 SDValue Root = Op.getOperand(0);
6203 SDValue Trmp = Op.getOperand(1); // trampoline
6204 SDValue FPtr = Op.getOperand(2); // nested function
6205 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6206 DebugLoc dl = Op.getDebugLoc();
6208 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6210 const X86InstrInfo *TII =
6211 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6213 if (Subtarget->is64Bit()) {
6214 SDValue OutChains[6];
6216 // Large code-model.
6218 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6219 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6221 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6222 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6224 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6226 // Load the pointer to the nested function into R11.
6227 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6228 SDValue Addr = Trmp;
6229 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6232 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6233 DAG.getConstant(2, MVT::i64));
6234 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6236 // Load the 'nest' parameter value into R10.
6237 // R10 is specified in X86CallingConv.td
6238 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6239 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6240 DAG.getConstant(10, MVT::i64));
6241 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6242 Addr, TrmpAddr, 10);
6244 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6245 DAG.getConstant(12, MVT::i64));
6246 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6248 // Jump to the nested function.
6249 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6250 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6251 DAG.getConstant(20, MVT::i64));
6252 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6253 Addr, TrmpAddr, 20);
6255 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6256 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6257 DAG.getConstant(22, MVT::i64));
6258 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6262 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6263 return DAG.getMergeValues(Ops, 2, dl);
6265 const Function *Func =
6266 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6267 unsigned CC = Func->getCallingConv();
6272 assert(0 && "Unsupported calling convention");
6273 case CallingConv::C:
6274 case CallingConv::X86_StdCall: {
6275 // Pass 'nest' parameter in ECX.
6276 // Must be kept in sync with X86CallingConv.td
6279 // Check that ECX wasn't needed by an 'inreg' parameter.
6280 const FunctionType *FTy = Func->getFunctionType();
6281 const AttrListPtr &Attrs = Func->getAttributes();
6283 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6284 unsigned InRegCount = 0;
6287 for (FunctionType::param_iterator I = FTy->param_begin(),
6288 E = FTy->param_end(); I != E; ++I, ++Idx)
6289 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6290 // FIXME: should only count parameters that are lowered to integers.
6291 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6293 if (InRegCount > 2) {
6294 cerr << "Nest register in use - reduce number of inreg parameters!\n";
6300 case CallingConv::X86_FastCall:
6301 case CallingConv::Fast:
6302 // Pass 'nest' parameter in EAX.
6303 // Must be kept in sync with X86CallingConv.td
6308 SDValue OutChains[4];
6311 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6312 DAG.getConstant(10, MVT::i32));
6313 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6315 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6316 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6317 OutChains[0] = DAG.getStore(Root, dl,
6318 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6321 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6322 DAG.getConstant(1, MVT::i32));
6323 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6325 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6326 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6327 DAG.getConstant(5, MVT::i32));
6328 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6329 TrmpAddr, 5, false, 1);
6331 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6332 DAG.getConstant(6, MVT::i32));
6333 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6336 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6337 return DAG.getMergeValues(Ops, 2, dl);
6341 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6343 The rounding mode is in bits 11:10 of FPSR, and has the following
6350 FLT_ROUNDS, on the other hand, expects the following:
6357 To perform the conversion, we do:
6358 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6361 MachineFunction &MF = DAG.getMachineFunction();
6362 const TargetMachine &TM = MF.getTarget();
6363 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6364 unsigned StackAlignment = TFI.getStackAlignment();
6365 MVT VT = Op.getValueType();
6366 DebugLoc dl = Op.getDebugLoc();
6368 // Save FP Control Word to stack slot
6369 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6370 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6372 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6373 DAG.getEntryNode(), StackSlot);
6375 // Load FP Control Word from stack slot
6376 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6378 // Transform as necessary
6380 DAG.getNode(ISD::SRL, dl, MVT::i16,
6381 DAG.getNode(ISD::AND, dl, MVT::i16,
6382 CWD, DAG.getConstant(0x800, MVT::i16)),
6383 DAG.getConstant(11, MVT::i8));
6385 DAG.getNode(ISD::SRL, dl, MVT::i16,
6386 DAG.getNode(ISD::AND, dl, MVT::i16,
6387 CWD, DAG.getConstant(0x400, MVT::i16)),
6388 DAG.getConstant(9, MVT::i8));
6391 DAG.getNode(ISD::AND, dl, MVT::i16,
6392 DAG.getNode(ISD::ADD, dl, MVT::i16,
6393 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6394 DAG.getConstant(1, MVT::i16)),
6395 DAG.getConstant(3, MVT::i16));
6398 return DAG.getNode((VT.getSizeInBits() < 16 ?
6399 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6402 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6403 MVT VT = Op.getValueType();
6405 unsigned NumBits = VT.getSizeInBits();
6406 DebugLoc dl = Op.getDebugLoc();
6408 Op = Op.getOperand(0);
6409 if (VT == MVT::i8) {
6410 // Zero extend to i32 since there is not an i8 bsr.
6412 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6415 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6416 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6417 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6419 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6420 SmallVector<SDValue, 4> Ops;
6422 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6423 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6424 Ops.push_back(Op.getValue(1));
6425 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6427 // Finally xor with NumBits-1.
6428 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6431 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6435 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6436 MVT VT = Op.getValueType();
6438 unsigned NumBits = VT.getSizeInBits();
6439 DebugLoc dl = Op.getDebugLoc();
6441 Op = Op.getOperand(0);
6442 if (VT == MVT::i8) {
6444 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6447 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6448 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6449 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6451 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6452 SmallVector<SDValue, 4> Ops;
6454 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6455 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6456 Ops.push_back(Op.getValue(1));
6457 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6460 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6464 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6465 MVT VT = Op.getValueType();
6466 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6467 DebugLoc dl = Op.getDebugLoc();
6469 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6470 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6471 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6472 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6473 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6475 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6476 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6477 // return AloBlo + AloBhi + AhiBlo;
6479 SDValue A = Op.getOperand(0);
6480 SDValue B = Op.getOperand(1);
6482 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6483 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6484 A, DAG.getConstant(32, MVT::i32));
6485 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6486 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6487 B, DAG.getConstant(32, MVT::i32));
6488 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6489 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6491 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6492 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6494 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6495 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6497 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6498 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6499 AloBhi, DAG.getConstant(32, MVT::i32));
6500 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6501 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6502 AhiBlo, DAG.getConstant(32, MVT::i32));
6503 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6504 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6509 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6510 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6511 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6512 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6513 // has only one use.
6514 SDNode *N = Op.getNode();
6515 SDValue LHS = N->getOperand(0);
6516 SDValue RHS = N->getOperand(1);
6517 unsigned BaseOp = 0;
6519 DebugLoc dl = Op.getDebugLoc();
6521 switch (Op.getOpcode()) {
6522 default: assert(0 && "Unknown ovf instruction!");
6524 // A subtract of one will be selected as a INC. Note that INC doesn't
6525 // set CF, so we can't do this for UADDO.
6526 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6527 if (C->getAPIntValue() == 1) {
6528 BaseOp = X86ISD::INC;
6532 BaseOp = X86ISD::ADD;
6536 BaseOp = X86ISD::ADD;
6540 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6541 // set CF, so we can't do this for USUBO.
6542 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6543 if (C->getAPIntValue() == 1) {
6544 BaseOp = X86ISD::DEC;
6548 BaseOp = X86ISD::SUB;
6552 BaseOp = X86ISD::SUB;
6556 BaseOp = X86ISD::SMUL;
6560 BaseOp = X86ISD::UMUL;
6565 // Also sets EFLAGS.
6566 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6567 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6570 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6571 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6573 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6577 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6578 MVT T = Op.getValueType();
6579 DebugLoc dl = Op.getDebugLoc();
6582 switch(T.getSimpleVT()) {
6584 assert(false && "Invalid value type!");
6585 case MVT::i8: Reg = X86::AL; size = 1; break;
6586 case MVT::i16: Reg = X86::AX; size = 2; break;
6587 case MVT::i32: Reg = X86::EAX; size = 4; break;
6589 assert(Subtarget->is64Bit() && "Node not type legal!");
6590 Reg = X86::RAX; size = 8;
6593 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6594 Op.getOperand(2), SDValue());
6595 SDValue Ops[] = { cpIn.getValue(0),
6598 DAG.getTargetConstant(size, MVT::i8),
6600 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6601 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6603 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6607 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6608 SelectionDAG &DAG) {
6609 assert(Subtarget->is64Bit() && "Result not type legalized?");
6610 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6611 SDValue TheChain = Op.getOperand(0);
6612 DebugLoc dl = Op.getDebugLoc();
6613 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6614 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6615 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6617 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6618 DAG.getConstant(32, MVT::i8));
6620 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6623 return DAG.getMergeValues(Ops, 2, dl);
6626 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6627 SDNode *Node = Op.getNode();
6628 DebugLoc dl = Node->getDebugLoc();
6629 MVT T = Node->getValueType(0);
6630 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6631 DAG.getConstant(0, T), Node->getOperand(2));
6632 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6633 cast<AtomicSDNode>(Node)->getMemoryVT(),
6634 Node->getOperand(0),
6635 Node->getOperand(1), negOp,
6636 cast<AtomicSDNode>(Node)->getSrcValue(),
6637 cast<AtomicSDNode>(Node)->getAlignment());
6640 /// LowerOperation - Provide custom lowering hooks for some operations.
6642 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6643 switch (Op.getOpcode()) {
6644 default: assert(0 && "Should not custom lower this!");
6645 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6646 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6647 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6648 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6649 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6650 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6651 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6652 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6653 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6654 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6655 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6656 case ISD::SHL_PARTS:
6657 case ISD::SRA_PARTS:
6658 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6659 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6660 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6661 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6662 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
6663 case ISD::FABS: return LowerFABS(Op, DAG);
6664 case ISD::FNEG: return LowerFNEG(Op, DAG);
6665 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6666 case ISD::SETCC: return LowerSETCC(Op, DAG);
6667 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6668 case ISD::SELECT: return LowerSELECT(Op, DAG);
6669 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6670 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6671 case ISD::CALL: return LowerCALL(Op, DAG);
6672 case ISD::RET: return LowerRET(Op, DAG);
6673 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6674 case ISD::VASTART: return LowerVASTART(Op, DAG);
6675 case ISD::VAARG: return LowerVAARG(Op, DAG);
6676 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6677 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6678 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6679 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6680 case ISD::FRAME_TO_ARGS_OFFSET:
6681 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6682 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6683 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6684 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6685 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6686 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6687 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6688 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6694 case ISD::UMULO: return LowerXALUO(Op, DAG);
6695 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6699 void X86TargetLowering::
6700 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6701 SelectionDAG &DAG, unsigned NewOp) {
6702 MVT T = Node->getValueType(0);
6703 DebugLoc dl = Node->getDebugLoc();
6704 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6706 SDValue Chain = Node->getOperand(0);
6707 SDValue In1 = Node->getOperand(1);
6708 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6709 Node->getOperand(2), DAG.getIntPtrConstant(0));
6710 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6711 Node->getOperand(2), DAG.getIntPtrConstant(1));
6712 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6713 // have a MemOperand. Pass the info through as a normal operand.
6714 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6715 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6716 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6717 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6718 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6719 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6720 Results.push_back(Result.getValue(2));
6723 /// ReplaceNodeResults - Replace a node with an illegal result type
6724 /// with a new node built out of custom code.
6725 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6726 SmallVectorImpl<SDValue>&Results,
6727 SelectionDAG &DAG) {
6728 DebugLoc dl = N->getDebugLoc();
6729 switch (N->getOpcode()) {
6731 assert(false && "Do not know how to custom type legalize this operation!");
6733 case ISD::FP_TO_SINT: {
6734 std::pair<SDValue,SDValue> Vals =
6735 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
6736 SDValue FIST = Vals.first, StackSlot = Vals.second;
6737 if (FIST.getNode() != 0) {
6738 MVT VT = N->getValueType(0);
6739 // Return a load from the stack slot.
6740 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6744 case ISD::READCYCLECOUNTER: {
6745 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6746 SDValue TheChain = N->getOperand(0);
6747 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6748 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6750 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6752 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6753 SDValue Ops[] = { eax, edx };
6754 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6755 Results.push_back(edx.getValue(1));
6758 case ISD::ATOMIC_CMP_SWAP: {
6759 MVT T = N->getValueType(0);
6760 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6761 SDValue cpInL, cpInH;
6762 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6763 DAG.getConstant(0, MVT::i32));
6764 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6765 DAG.getConstant(1, MVT::i32));
6766 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6767 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6769 SDValue swapInL, swapInH;
6770 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6771 DAG.getConstant(0, MVT::i32));
6772 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6773 DAG.getConstant(1, MVT::i32));
6774 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6776 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6777 swapInL.getValue(1));
6778 SDValue Ops[] = { swapInH.getValue(0),
6780 swapInH.getValue(1) };
6781 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6782 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6783 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6784 MVT::i32, Result.getValue(1));
6785 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6786 MVT::i32, cpOutL.getValue(2));
6787 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6788 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6789 Results.push_back(cpOutH.getValue(1));
6792 case ISD::ATOMIC_LOAD_ADD:
6793 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6795 case ISD::ATOMIC_LOAD_AND:
6796 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6798 case ISD::ATOMIC_LOAD_NAND:
6799 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6801 case ISD::ATOMIC_LOAD_OR:
6802 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6804 case ISD::ATOMIC_LOAD_SUB:
6805 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6807 case ISD::ATOMIC_LOAD_XOR:
6808 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6810 case ISD::ATOMIC_SWAP:
6811 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6816 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6818 default: return NULL;
6819 case X86ISD::BSF: return "X86ISD::BSF";
6820 case X86ISD::BSR: return "X86ISD::BSR";
6821 case X86ISD::SHLD: return "X86ISD::SHLD";
6822 case X86ISD::SHRD: return "X86ISD::SHRD";
6823 case X86ISD::FAND: return "X86ISD::FAND";
6824 case X86ISD::FOR: return "X86ISD::FOR";
6825 case X86ISD::FXOR: return "X86ISD::FXOR";
6826 case X86ISD::FSRL: return "X86ISD::FSRL";
6827 case X86ISD::FILD: return "X86ISD::FILD";
6828 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6829 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6830 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6831 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6832 case X86ISD::FLD: return "X86ISD::FLD";
6833 case X86ISD::FST: return "X86ISD::FST";
6834 case X86ISD::CALL: return "X86ISD::CALL";
6835 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6836 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6837 case X86ISD::BT: return "X86ISD::BT";
6838 case X86ISD::CMP: return "X86ISD::CMP";
6839 case X86ISD::COMI: return "X86ISD::COMI";
6840 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6841 case X86ISD::SETCC: return "X86ISD::SETCC";
6842 case X86ISD::CMOV: return "X86ISD::CMOV";
6843 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6844 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6845 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6846 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6847 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6848 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6849 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
6850 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6851 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6852 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6853 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6854 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6855 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
6856 case X86ISD::FMAX: return "X86ISD::FMAX";
6857 case X86ISD::FMIN: return "X86ISD::FMIN";
6858 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6859 case X86ISD::FRCP: return "X86ISD::FRCP";
6860 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6861 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
6862 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6863 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6864 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6865 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6866 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6867 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
6868 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
6869 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
6870 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
6871 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
6872 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
6873 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6874 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6875 case X86ISD::VSHL: return "X86ISD::VSHL";
6876 case X86ISD::VSRL: return "X86ISD::VSRL";
6877 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6878 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6879 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6880 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6881 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6882 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6883 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6884 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6885 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6886 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6887 case X86ISD::ADD: return "X86ISD::ADD";
6888 case X86ISD::SUB: return "X86ISD::SUB";
6889 case X86ISD::SMUL: return "X86ISD::SMUL";
6890 case X86ISD::UMUL: return "X86ISD::UMUL";
6891 case X86ISD::INC: return "X86ISD::INC";
6892 case X86ISD::DEC: return "X86ISD::DEC";
6893 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
6897 // isLegalAddressingMode - Return true if the addressing mode represented
6898 // by AM is legal for this target, for a load/store of the specified type.
6899 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6900 const Type *Ty) const {
6901 // X86 supports extremely general addressing modes.
6903 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6904 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6908 // We can only fold this if we don't need an extra load.
6909 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6911 // If BaseGV requires a register, we cannot also have a BaseReg.
6912 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
6916 // X86-64 only supports addr of globals in small code model.
6917 if (Subtarget->is64Bit()) {
6918 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6920 // If lower 4G is not available, then we must use rip-relative addressing.
6921 if (AM.BaseOffs || AM.Scale > 1)
6932 // These scales always work.
6937 // These scales are formed with basereg+scalereg. Only accept if there is
6942 default: // Other stuff never works.
6950 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6951 if (!Ty1->isInteger() || !Ty2->isInteger())
6953 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6954 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6955 if (NumBits1 <= NumBits2)
6957 return Subtarget->is64Bit() || NumBits1 < 64;
6960 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6961 if (!VT1.isInteger() || !VT2.isInteger())
6963 unsigned NumBits1 = VT1.getSizeInBits();
6964 unsigned NumBits2 = VT2.getSizeInBits();
6965 if (NumBits1 <= NumBits2)
6967 return Subtarget->is64Bit() || NumBits1 < 64;
6970 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
6971 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
6972 return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
6975 bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
6976 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
6977 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
6980 bool X86TargetLowering::isNarrowingProfitable(MVT VT1, MVT VT2) const {
6981 // i16 instructions are longer (0x66 prefix) and potentially slower.
6982 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
6985 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6986 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6987 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6988 /// are assumed to be legal.
6990 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6992 // Only do shuffles on 128-bit vector types for now.
6993 if (VT.getSizeInBits() == 64)
6996 // FIXME: pshufb, blends, palignr, shifts.
6997 return (VT.getVectorNumElements() == 2 ||
6998 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
6999 isMOVLMask(M, VT) ||
7000 isSHUFPMask(M, VT) ||
7001 isPSHUFDMask(M, VT) ||
7002 isPSHUFHWMask(M, VT) ||
7003 isPSHUFLWMask(M, VT) ||
7004 isUNPCKLMask(M, VT) ||
7005 isUNPCKHMask(M, VT) ||
7006 isUNPCKL_v_undef_Mask(M, VT) ||
7007 isUNPCKH_v_undef_Mask(M, VT));
7011 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7013 unsigned NumElts = VT.getVectorNumElements();
7014 // FIXME: This collection of masks seems suspect.
7017 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7018 return (isMOVLMask(Mask, VT) ||
7019 isCommutedMOVLMask(Mask, VT, true) ||
7020 isSHUFPMask(Mask, VT) ||
7021 isCommutedSHUFPMask(Mask, VT));
7026 //===----------------------------------------------------------------------===//
7027 // X86 Scheduler Hooks
7028 //===----------------------------------------------------------------------===//
7030 // private utility function
7032 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7033 MachineBasicBlock *MBB,
7041 TargetRegisterClass *RC,
7042 bool invSrc) const {
7043 // For the atomic bitwise operator, we generate
7046 // ld t1 = [bitinstr.addr]
7047 // op t2 = t1, [bitinstr.val]
7049 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7051 // fallthrough -->nextMBB
7052 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7053 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7054 MachineFunction::iterator MBBIter = MBB;
7057 /// First build the CFG
7058 MachineFunction *F = MBB->getParent();
7059 MachineBasicBlock *thisMBB = MBB;
7060 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7061 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7062 F->insert(MBBIter, newMBB);
7063 F->insert(MBBIter, nextMBB);
7065 // Move all successors to thisMBB to nextMBB
7066 nextMBB->transferSuccessors(thisMBB);
7068 // Update thisMBB to fall through to newMBB
7069 thisMBB->addSuccessor(newMBB);
7071 // newMBB jumps to itself and fall through to nextMBB
7072 newMBB->addSuccessor(nextMBB);
7073 newMBB->addSuccessor(newMBB);
7075 // Insert instructions into newMBB based on incoming instruction
7076 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7077 "unexpected number of operands");
7078 DebugLoc dl = bInstr->getDebugLoc();
7079 MachineOperand& destOper = bInstr->getOperand(0);
7080 MachineOperand* argOpers[2 + X86AddrNumOperands];
7081 int numArgs = bInstr->getNumOperands() - 1;
7082 for (int i=0; i < numArgs; ++i)
7083 argOpers[i] = &bInstr->getOperand(i+1);
7085 // x86 address has 4 operands: base, index, scale, and displacement
7086 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7087 int valArgIndx = lastAddrIndx + 1;
7089 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7090 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7091 for (int i=0; i <= lastAddrIndx; ++i)
7092 (*MIB).addOperand(*argOpers[i]);
7094 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7096 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7101 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7102 assert((argOpers[valArgIndx]->isReg() ||
7103 argOpers[valArgIndx]->isImm()) &&
7105 if (argOpers[valArgIndx]->isReg())
7106 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7108 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7110 (*MIB).addOperand(*argOpers[valArgIndx]);
7112 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7115 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7116 for (int i=0; i <= lastAddrIndx; ++i)
7117 (*MIB).addOperand(*argOpers[i]);
7119 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7120 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7122 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7126 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7128 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7132 // private utility function: 64 bit atomics on 32 bit host.
7134 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7135 MachineBasicBlock *MBB,
7140 bool invSrc) const {
7141 // For the atomic bitwise operator, we generate
7142 // thisMBB (instructions are in pairs, except cmpxchg8b)
7143 // ld t1,t2 = [bitinstr.addr]
7145 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7146 // op t5, t6 <- out1, out2, [bitinstr.val]
7147 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7148 // mov ECX, EBX <- t5, t6
7149 // mov EAX, EDX <- t1, t2
7150 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7151 // mov t3, t4 <- EAX, EDX
7153 // result in out1, out2
7154 // fallthrough -->nextMBB
7156 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7157 const unsigned LoadOpc = X86::MOV32rm;
7158 const unsigned copyOpc = X86::MOV32rr;
7159 const unsigned NotOpc = X86::NOT32r;
7160 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7161 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7162 MachineFunction::iterator MBBIter = MBB;
7165 /// First build the CFG
7166 MachineFunction *F = MBB->getParent();
7167 MachineBasicBlock *thisMBB = MBB;
7168 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7169 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7170 F->insert(MBBIter, newMBB);
7171 F->insert(MBBIter, nextMBB);
7173 // Move all successors to thisMBB to nextMBB
7174 nextMBB->transferSuccessors(thisMBB);
7176 // Update thisMBB to fall through to newMBB
7177 thisMBB->addSuccessor(newMBB);
7179 // newMBB jumps to itself and fall through to nextMBB
7180 newMBB->addSuccessor(nextMBB);
7181 newMBB->addSuccessor(newMBB);
7183 DebugLoc dl = bInstr->getDebugLoc();
7184 // Insert instructions into newMBB based on incoming instruction
7185 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7186 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7187 "unexpected number of operands");
7188 MachineOperand& dest1Oper = bInstr->getOperand(0);
7189 MachineOperand& dest2Oper = bInstr->getOperand(1);
7190 MachineOperand* argOpers[2 + X86AddrNumOperands];
7191 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7192 argOpers[i] = &bInstr->getOperand(i+2);
7194 // x86 address has 4 operands: base, index, scale, and displacement
7195 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7197 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7198 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7199 for (int i=0; i <= lastAddrIndx; ++i)
7200 (*MIB).addOperand(*argOpers[i]);
7201 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7202 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7203 // add 4 to displacement.
7204 for (int i=0; i <= lastAddrIndx-2; ++i)
7205 (*MIB).addOperand(*argOpers[i]);
7206 MachineOperand newOp3 = *(argOpers[3]);
7208 newOp3.setImm(newOp3.getImm()+4);
7210 newOp3.setOffset(newOp3.getOffset()+4);
7211 (*MIB).addOperand(newOp3);
7212 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7214 // t3/4 are defined later, at the bottom of the loop
7215 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7216 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7217 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7218 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7219 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7220 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7222 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7223 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7225 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7226 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7232 int valArgIndx = lastAddrIndx + 1;
7233 assert((argOpers[valArgIndx]->isReg() ||
7234 argOpers[valArgIndx]->isImm()) &&
7236 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7237 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7238 if (argOpers[valArgIndx]->isReg())
7239 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7241 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7242 if (regOpcL != X86::MOV32rr)
7244 (*MIB).addOperand(*argOpers[valArgIndx]);
7245 assert(argOpers[valArgIndx + 1]->isReg() ==
7246 argOpers[valArgIndx]->isReg());
7247 assert(argOpers[valArgIndx + 1]->isImm() ==
7248 argOpers[valArgIndx]->isImm());
7249 if (argOpers[valArgIndx + 1]->isReg())
7250 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7252 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7253 if (regOpcH != X86::MOV32rr)
7255 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7257 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7259 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7262 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7264 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7267 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7268 for (int i=0; i <= lastAddrIndx; ++i)
7269 (*MIB).addOperand(*argOpers[i]);
7271 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7272 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7274 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7275 MIB.addReg(X86::EAX);
7276 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7277 MIB.addReg(X86::EDX);
7280 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7282 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7286 // private utility function
7288 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7289 MachineBasicBlock *MBB,
7290 unsigned cmovOpc) const {
7291 // For the atomic min/max operator, we generate
7294 // ld t1 = [min/max.addr]
7295 // mov t2 = [min/max.val]
7297 // cmov[cond] t2 = t1
7299 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7301 // fallthrough -->nextMBB
7303 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7304 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7305 MachineFunction::iterator MBBIter = MBB;
7308 /// First build the CFG
7309 MachineFunction *F = MBB->getParent();
7310 MachineBasicBlock *thisMBB = MBB;
7311 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7312 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7313 F->insert(MBBIter, newMBB);
7314 F->insert(MBBIter, nextMBB);
7316 // Move all successors to thisMBB to nextMBB
7317 nextMBB->transferSuccessors(thisMBB);
7319 // Update thisMBB to fall through to newMBB
7320 thisMBB->addSuccessor(newMBB);
7322 // newMBB jumps to newMBB and fall through to nextMBB
7323 newMBB->addSuccessor(nextMBB);
7324 newMBB->addSuccessor(newMBB);
7326 DebugLoc dl = mInstr->getDebugLoc();
7327 // Insert instructions into newMBB based on incoming instruction
7328 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7329 "unexpected number of operands");
7330 MachineOperand& destOper = mInstr->getOperand(0);
7331 MachineOperand* argOpers[2 + X86AddrNumOperands];
7332 int numArgs = mInstr->getNumOperands() - 1;
7333 for (int i=0; i < numArgs; ++i)
7334 argOpers[i] = &mInstr->getOperand(i+1);
7336 // x86 address has 4 operands: base, index, scale, and displacement
7337 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7338 int valArgIndx = lastAddrIndx + 1;
7340 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7341 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7342 for (int i=0; i <= lastAddrIndx; ++i)
7343 (*MIB).addOperand(*argOpers[i]);
7345 // We only support register and immediate values
7346 assert((argOpers[valArgIndx]->isReg() ||
7347 argOpers[valArgIndx]->isImm()) &&
7350 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7351 if (argOpers[valArgIndx]->isReg())
7352 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7354 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7355 (*MIB).addOperand(*argOpers[valArgIndx]);
7357 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7360 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7365 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7366 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7370 // Cmp and exchange if none has modified the memory location
7371 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7372 for (int i=0; i <= lastAddrIndx; ++i)
7373 (*MIB).addOperand(*argOpers[i]);
7375 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7376 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7378 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7379 MIB.addReg(X86::EAX);
7382 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7384 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7390 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7391 MachineBasicBlock *BB) const {
7392 DebugLoc dl = MI->getDebugLoc();
7393 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7394 switch (MI->getOpcode()) {
7395 default: assert(false && "Unexpected instr type to insert");
7396 case X86::CMOV_V1I64:
7397 case X86::CMOV_FR32:
7398 case X86::CMOV_FR64:
7399 case X86::CMOV_V4F32:
7400 case X86::CMOV_V2F64:
7401 case X86::CMOV_V2I64: {
7402 // To "insert" a SELECT_CC instruction, we actually have to insert the
7403 // diamond control-flow pattern. The incoming instruction knows the
7404 // destination vreg to set, the condition code register to branch on, the
7405 // true/false values to select between, and a branch opcode to use.
7406 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7407 MachineFunction::iterator It = BB;
7413 // cmpTY ccX, r1, r2
7415 // fallthrough --> copy0MBB
7416 MachineBasicBlock *thisMBB = BB;
7417 MachineFunction *F = BB->getParent();
7418 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7419 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7421 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7422 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7423 F->insert(It, copy0MBB);
7424 F->insert(It, sinkMBB);
7425 // Update machine-CFG edges by transferring all successors of the current
7426 // block to the new block which will contain the Phi node for the select.
7427 sinkMBB->transferSuccessors(BB);
7429 // Add the true and fallthrough blocks as its successors.
7430 BB->addSuccessor(copy0MBB);
7431 BB->addSuccessor(sinkMBB);
7434 // %FalseValue = ...
7435 // # fallthrough to sinkMBB
7438 // Update machine-CFG edges
7439 BB->addSuccessor(sinkMBB);
7442 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7445 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7446 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7447 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7449 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7453 case X86::FP32_TO_INT16_IN_MEM:
7454 case X86::FP32_TO_INT32_IN_MEM:
7455 case X86::FP32_TO_INT64_IN_MEM:
7456 case X86::FP64_TO_INT16_IN_MEM:
7457 case X86::FP64_TO_INT32_IN_MEM:
7458 case X86::FP64_TO_INT64_IN_MEM:
7459 case X86::FP80_TO_INT16_IN_MEM:
7460 case X86::FP80_TO_INT32_IN_MEM:
7461 case X86::FP80_TO_INT64_IN_MEM: {
7462 // Change the floating point control register to use "round towards zero"
7463 // mode when truncating to an integer value.
7464 MachineFunction *F = BB->getParent();
7465 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7466 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7468 // Load the old value of the high byte of the control word...
7470 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7471 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7474 // Set the high part to be round to zero...
7475 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7478 // Reload the modified control word now...
7479 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7481 // Restore the memory image of control word to original value
7482 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7485 // Get the X86 opcode to use.
7487 switch (MI->getOpcode()) {
7488 default: assert(0 && "illegal opcode!");
7489 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7490 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7491 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7492 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7493 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7494 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7495 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7496 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7497 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7501 MachineOperand &Op = MI->getOperand(0);
7503 AM.BaseType = X86AddressMode::RegBase;
7504 AM.Base.Reg = Op.getReg();
7506 AM.BaseType = X86AddressMode::FrameIndexBase;
7507 AM.Base.FrameIndex = Op.getIndex();
7509 Op = MI->getOperand(1);
7511 AM.Scale = Op.getImm();
7512 Op = MI->getOperand(2);
7514 AM.IndexReg = Op.getImm();
7515 Op = MI->getOperand(3);
7516 if (Op.isGlobal()) {
7517 AM.GV = Op.getGlobal();
7519 AM.Disp = Op.getImm();
7521 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7522 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7524 // Reload the original control word now.
7525 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7527 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7530 case X86::ATOMAND32:
7531 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7532 X86::AND32ri, X86::MOV32rm,
7533 X86::LCMPXCHG32, X86::MOV32rr,
7534 X86::NOT32r, X86::EAX,
7535 X86::GR32RegisterClass);
7537 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7538 X86::OR32ri, X86::MOV32rm,
7539 X86::LCMPXCHG32, X86::MOV32rr,
7540 X86::NOT32r, X86::EAX,
7541 X86::GR32RegisterClass);
7542 case X86::ATOMXOR32:
7543 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7544 X86::XOR32ri, X86::MOV32rm,
7545 X86::LCMPXCHG32, X86::MOV32rr,
7546 X86::NOT32r, X86::EAX,
7547 X86::GR32RegisterClass);
7548 case X86::ATOMNAND32:
7549 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7550 X86::AND32ri, X86::MOV32rm,
7551 X86::LCMPXCHG32, X86::MOV32rr,
7552 X86::NOT32r, X86::EAX,
7553 X86::GR32RegisterClass, true);
7554 case X86::ATOMMIN32:
7555 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7556 case X86::ATOMMAX32:
7557 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7558 case X86::ATOMUMIN32:
7559 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7560 case X86::ATOMUMAX32:
7561 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7563 case X86::ATOMAND16:
7564 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7565 X86::AND16ri, X86::MOV16rm,
7566 X86::LCMPXCHG16, X86::MOV16rr,
7567 X86::NOT16r, X86::AX,
7568 X86::GR16RegisterClass);
7570 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7571 X86::OR16ri, X86::MOV16rm,
7572 X86::LCMPXCHG16, X86::MOV16rr,
7573 X86::NOT16r, X86::AX,
7574 X86::GR16RegisterClass);
7575 case X86::ATOMXOR16:
7576 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7577 X86::XOR16ri, X86::MOV16rm,
7578 X86::LCMPXCHG16, X86::MOV16rr,
7579 X86::NOT16r, X86::AX,
7580 X86::GR16RegisterClass);
7581 case X86::ATOMNAND16:
7582 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7583 X86::AND16ri, X86::MOV16rm,
7584 X86::LCMPXCHG16, X86::MOV16rr,
7585 X86::NOT16r, X86::AX,
7586 X86::GR16RegisterClass, true);
7587 case X86::ATOMMIN16:
7588 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7589 case X86::ATOMMAX16:
7590 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7591 case X86::ATOMUMIN16:
7592 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7593 case X86::ATOMUMAX16:
7594 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7597 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7598 X86::AND8ri, X86::MOV8rm,
7599 X86::LCMPXCHG8, X86::MOV8rr,
7600 X86::NOT8r, X86::AL,
7601 X86::GR8RegisterClass);
7603 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7604 X86::OR8ri, X86::MOV8rm,
7605 X86::LCMPXCHG8, X86::MOV8rr,
7606 X86::NOT8r, X86::AL,
7607 X86::GR8RegisterClass);
7609 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7610 X86::XOR8ri, X86::MOV8rm,
7611 X86::LCMPXCHG8, X86::MOV8rr,
7612 X86::NOT8r, X86::AL,
7613 X86::GR8RegisterClass);
7614 case X86::ATOMNAND8:
7615 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7616 X86::AND8ri, X86::MOV8rm,
7617 X86::LCMPXCHG8, X86::MOV8rr,
7618 X86::NOT8r, X86::AL,
7619 X86::GR8RegisterClass, true);
7620 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7621 // This group is for 64-bit host.
7622 case X86::ATOMAND64:
7623 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7624 X86::AND64ri32, X86::MOV64rm,
7625 X86::LCMPXCHG64, X86::MOV64rr,
7626 X86::NOT64r, X86::RAX,
7627 X86::GR64RegisterClass);
7629 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7630 X86::OR64ri32, X86::MOV64rm,
7631 X86::LCMPXCHG64, X86::MOV64rr,
7632 X86::NOT64r, X86::RAX,
7633 X86::GR64RegisterClass);
7634 case X86::ATOMXOR64:
7635 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7636 X86::XOR64ri32, X86::MOV64rm,
7637 X86::LCMPXCHG64, X86::MOV64rr,
7638 X86::NOT64r, X86::RAX,
7639 X86::GR64RegisterClass);
7640 case X86::ATOMNAND64:
7641 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7642 X86::AND64ri32, X86::MOV64rm,
7643 X86::LCMPXCHG64, X86::MOV64rr,
7644 X86::NOT64r, X86::RAX,
7645 X86::GR64RegisterClass, true);
7646 case X86::ATOMMIN64:
7647 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7648 case X86::ATOMMAX64:
7649 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7650 case X86::ATOMUMIN64:
7651 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7652 case X86::ATOMUMAX64:
7653 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7655 // This group does 64-bit operations on a 32-bit host.
7656 case X86::ATOMAND6432:
7657 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7658 X86::AND32rr, X86::AND32rr,
7659 X86::AND32ri, X86::AND32ri,
7661 case X86::ATOMOR6432:
7662 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7663 X86::OR32rr, X86::OR32rr,
7664 X86::OR32ri, X86::OR32ri,
7666 case X86::ATOMXOR6432:
7667 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7668 X86::XOR32rr, X86::XOR32rr,
7669 X86::XOR32ri, X86::XOR32ri,
7671 case X86::ATOMNAND6432:
7672 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7673 X86::AND32rr, X86::AND32rr,
7674 X86::AND32ri, X86::AND32ri,
7676 case X86::ATOMADD6432:
7677 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7678 X86::ADD32rr, X86::ADC32rr,
7679 X86::ADD32ri, X86::ADC32ri,
7681 case X86::ATOMSUB6432:
7682 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7683 X86::SUB32rr, X86::SBB32rr,
7684 X86::SUB32ri, X86::SBB32ri,
7686 case X86::ATOMSWAP6432:
7687 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7688 X86::MOV32rr, X86::MOV32rr,
7689 X86::MOV32ri, X86::MOV32ri,
7694 //===----------------------------------------------------------------------===//
7695 // X86 Optimization Hooks
7696 //===----------------------------------------------------------------------===//
7698 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7702 const SelectionDAG &DAG,
7703 unsigned Depth) const {
7704 unsigned Opc = Op.getOpcode();
7705 assert((Opc >= ISD::BUILTIN_OP_END ||
7706 Opc == ISD::INTRINSIC_WO_CHAIN ||
7707 Opc == ISD::INTRINSIC_W_CHAIN ||
7708 Opc == ISD::INTRINSIC_VOID) &&
7709 "Should use MaskedValueIsZero if you don't know whether Op"
7710 " is a target node!");
7712 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7721 // These nodes' second result is a boolean.
7722 if (Op.getResNo() == 0)
7726 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7727 Mask.getBitWidth() - 1);
7732 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7733 /// node is a GlobalAddress + offset.
7734 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7735 GlobalValue* &GA, int64_t &Offset) const{
7736 if (N->getOpcode() == X86ISD::Wrapper) {
7737 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7738 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7739 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7743 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7746 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7747 const TargetLowering &TLI) {
7750 if (TLI.isGAPlusOffset(Base, GV, Offset))
7751 return (GV->getAlignment() >= N && (Offset % N) == 0);
7752 // DAG combine handles the stack object case.
7756 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
7757 MVT EVT, LoadSDNode *&LDBase,
7758 unsigned &LastLoadedElt,
7759 SelectionDAG &DAG, MachineFrameInfo *MFI,
7760 const TargetLowering &TLI) {
7762 LastLoadedElt = -1U;
7763 for (unsigned i = 0; i < NumElems; ++i) {
7764 if (N->getMaskElt(i) < 0) {
7770 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7771 if (!Elt.getNode() ||
7772 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7775 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
7777 LDBase = cast<LoadSDNode>(Elt.getNode());
7781 if (Elt.getOpcode() == ISD::UNDEF)
7784 LoadSDNode *LD = cast<LoadSDNode>(Elt);
7785 if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
7792 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7793 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7794 /// if the load addresses are consecutive, non-overlapping, and in the right
7795 /// order. In the case of v2i64, it will see if it can rewrite the
7796 /// shuffle to be an appropriate build vector so it can take advantage of
7797 // performBuildVectorCombine.
7798 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7799 const TargetLowering &TLI) {
7800 DebugLoc dl = N->getDebugLoc();
7801 MVT VT = N->getValueType(0);
7802 MVT EVT = VT.getVectorElementType();
7803 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7804 unsigned NumElems = VT.getVectorNumElements();
7806 if (VT.getSizeInBits() != 128)
7809 // Try to combine a vector_shuffle into a 128-bit load.
7810 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7811 LoadSDNode *LD = NULL;
7812 unsigned LastLoadedElt;
7813 if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
7817 if (LastLoadedElt == NumElems - 1) {
7818 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
7819 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7820 LD->getSrcValue(), LD->getSrcValueOffset(),
7822 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7823 LD->getSrcValue(), LD->getSrcValueOffset(),
7824 LD->isVolatile(), LD->getAlignment());
7825 } else if (NumElems == 4 && LastLoadedElt == 1) {
7826 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
7827 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7828 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
7829 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
7834 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7835 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7836 const X86Subtarget *Subtarget) {
7837 DebugLoc DL = N->getDebugLoc();
7838 SDValue Cond = N->getOperand(0);
7839 // Get the LHS/RHS of the select.
7840 SDValue LHS = N->getOperand(1);
7841 SDValue RHS = N->getOperand(2);
7843 // If we have SSE[12] support, try to form min/max nodes.
7844 if (Subtarget->hasSSE2() &&
7845 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
7846 Cond.getOpcode() == ISD::SETCC) {
7847 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7849 unsigned Opcode = 0;
7850 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7853 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
7856 if (!UnsafeFPMath) break;
7858 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
7860 Opcode = X86ISD::FMIN;
7863 case ISD::SETOGT: // (X > Y) ? X : Y -> max
7866 if (!UnsafeFPMath) break;
7868 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
7870 Opcode = X86ISD::FMAX;
7873 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
7876 case ISD::SETOGT: // (X > Y) ? Y : X -> min
7879 if (!UnsafeFPMath) break;
7881 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
7883 Opcode = X86ISD::FMIN;
7886 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
7889 if (!UnsafeFPMath) break;
7891 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
7893 Opcode = X86ISD::FMAX;
7899 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
7902 // If this is a select between two integer constants, try to do some
7904 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
7905 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
7906 // Don't do this for crazy integer types.
7907 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
7908 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
7909 // so that TrueC (the true value) is larger than FalseC.
7910 bool NeedsCondInvert = false;
7912 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
7913 // Efficiently invertible.
7914 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
7915 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
7916 isa<ConstantSDNode>(Cond.getOperand(1))))) {
7917 NeedsCondInvert = true;
7918 std::swap(TrueC, FalseC);
7921 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
7922 if (FalseC->getAPIntValue() == 0 &&
7923 TrueC->getAPIntValue().isPowerOf2()) {
7924 if (NeedsCondInvert) // Invert the condition if needed.
7925 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7926 DAG.getConstant(1, Cond.getValueType()));
7928 // Zero extend the condition if needed.
7929 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
7931 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
7932 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
7933 DAG.getConstant(ShAmt, MVT::i8));
7936 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
7937 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
7938 if (NeedsCondInvert) // Invert the condition if needed.
7939 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7940 DAG.getConstant(1, Cond.getValueType()));
7942 // Zero extend the condition if needed.
7943 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
7944 FalseC->getValueType(0), Cond);
7945 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
7946 SDValue(FalseC, 0));
7949 // Optimize cases that will turn into an LEA instruction. This requires
7950 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
7951 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
7952 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
7953 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
7955 bool isFastMultiplier = false;
7957 switch ((unsigned char)Diff) {
7959 case 1: // result = add base, cond
7960 case 2: // result = lea base( , cond*2)
7961 case 3: // result = lea base(cond, cond*2)
7962 case 4: // result = lea base( , cond*4)
7963 case 5: // result = lea base(cond, cond*4)
7964 case 8: // result = lea base( , cond*8)
7965 case 9: // result = lea base(cond, cond*8)
7966 isFastMultiplier = true;
7971 if (isFastMultiplier) {
7972 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
7973 if (NeedsCondInvert) // Invert the condition if needed.
7974 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7975 DAG.getConstant(1, Cond.getValueType()));
7977 // Zero extend the condition if needed.
7978 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
7980 // Scale the condition by the difference.
7982 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
7983 DAG.getConstant(Diff, Cond.getValueType()));
7985 // Add the base if non-zero.
7986 if (FalseC->getAPIntValue() != 0)
7987 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
7988 SDValue(FalseC, 0));
7998 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
7999 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8000 TargetLowering::DAGCombinerInfo &DCI) {
8001 DebugLoc DL = N->getDebugLoc();
8003 // If the flag operand isn't dead, don't touch this CMOV.
8004 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8007 // If this is a select between two integer constants, try to do some
8008 // optimizations. Note that the operands are ordered the opposite of SELECT
8010 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8011 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8012 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8013 // larger than FalseC (the false value).
8014 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8016 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8017 CC = X86::GetOppositeBranchCondition(CC);
8018 std::swap(TrueC, FalseC);
8021 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8022 // This is efficient for any integer data type (including i8/i16) and
8024 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8025 SDValue Cond = N->getOperand(3);
8026 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8027 DAG.getConstant(CC, MVT::i8), Cond);
8029 // Zero extend the condition if needed.
8030 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8032 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8033 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8034 DAG.getConstant(ShAmt, MVT::i8));
8035 if (N->getNumValues() == 2) // Dead flag value?
8036 return DCI.CombineTo(N, Cond, SDValue());
8040 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8041 // for any integer data type, including i8/i16.
8042 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8043 SDValue Cond = N->getOperand(3);
8044 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8045 DAG.getConstant(CC, MVT::i8), Cond);
8047 // Zero extend the condition if needed.
8048 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8049 FalseC->getValueType(0), Cond);
8050 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8051 SDValue(FalseC, 0));
8053 if (N->getNumValues() == 2) // Dead flag value?
8054 return DCI.CombineTo(N, Cond, SDValue());
8058 // Optimize cases that will turn into an LEA instruction. This requires
8059 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8060 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8061 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8062 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8064 bool isFastMultiplier = false;
8066 switch ((unsigned char)Diff) {
8068 case 1: // result = add base, cond
8069 case 2: // result = lea base( , cond*2)
8070 case 3: // result = lea base(cond, cond*2)
8071 case 4: // result = lea base( , cond*4)
8072 case 5: // result = lea base(cond, cond*4)
8073 case 8: // result = lea base( , cond*8)
8074 case 9: // result = lea base(cond, cond*8)
8075 isFastMultiplier = true;
8080 if (isFastMultiplier) {
8081 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8082 SDValue Cond = N->getOperand(3);
8083 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8084 DAG.getConstant(CC, MVT::i8), Cond);
8085 // Zero extend the condition if needed.
8086 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8088 // Scale the condition by the difference.
8090 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8091 DAG.getConstant(Diff, Cond.getValueType()));
8093 // Add the base if non-zero.
8094 if (FalseC->getAPIntValue() != 0)
8095 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8096 SDValue(FalseC, 0));
8097 if (N->getNumValues() == 2) // Dead flag value?
8098 return DCI.CombineTo(N, Cond, SDValue());
8108 /// PerformMulCombine - Optimize a single multiply with constant into two
8109 /// in order to implement it with two cheaper instructions, e.g.
8110 /// LEA + SHL, LEA + LEA.
8111 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8112 TargetLowering::DAGCombinerInfo &DCI) {
8113 if (DAG.getMachineFunction().
8114 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8117 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8120 MVT VT = N->getValueType(0);
8124 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8127 uint64_t MulAmt = C->getZExtValue();
8128 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8131 uint64_t MulAmt1 = 0;
8132 uint64_t MulAmt2 = 0;
8133 if ((MulAmt % 9) == 0) {
8135 MulAmt2 = MulAmt / 9;
8136 } else if ((MulAmt % 5) == 0) {
8138 MulAmt2 = MulAmt / 5;
8139 } else if ((MulAmt % 3) == 0) {
8141 MulAmt2 = MulAmt / 3;
8144 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8145 DebugLoc DL = N->getDebugLoc();
8147 if (isPowerOf2_64(MulAmt2) &&
8148 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8149 // If second multiplifer is pow2, issue it first. We want the multiply by
8150 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8152 std::swap(MulAmt1, MulAmt2);
8155 if (isPowerOf2_64(MulAmt1))
8156 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8157 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8159 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8160 DAG.getConstant(MulAmt1, VT));
8162 if (isPowerOf2_64(MulAmt2))
8163 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8164 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8166 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8167 DAG.getConstant(MulAmt2, VT));
8169 // Do not add new nodes to DAG combiner worklist.
8170 DCI.CombineTo(N, NewMul, false);
8176 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8178 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8179 const X86Subtarget *Subtarget) {
8180 // On X86 with SSE2 support, we can transform this to a vector shift if
8181 // all elements are shifted by the same amount. We can't do this in legalize
8182 // because the a constant vector is typically transformed to a constant pool
8183 // so we have no knowledge of the shift amount.
8184 if (!Subtarget->hasSSE2())
8187 MVT VT = N->getValueType(0);
8188 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8191 SDValue ShAmtOp = N->getOperand(1);
8192 MVT EltVT = VT.getVectorElementType();
8193 DebugLoc DL = N->getDebugLoc();
8195 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8196 unsigned NumElts = VT.getVectorNumElements();
8198 for (; i != NumElts; ++i) {
8199 SDValue Arg = ShAmtOp.getOperand(i);
8200 if (Arg.getOpcode() == ISD::UNDEF) continue;
8204 for (; i != NumElts; ++i) {
8205 SDValue Arg = ShAmtOp.getOperand(i);
8206 if (Arg.getOpcode() == ISD::UNDEF) continue;
8207 if (Arg != BaseShAmt) {
8211 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8212 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8213 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8214 DAG.getIntPtrConstant(0));
8218 if (EltVT.bitsGT(MVT::i32))
8219 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8220 else if (EltVT.bitsLT(MVT::i32))
8221 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8223 // The shift amount is identical so we can do a vector shift.
8224 SDValue ValOp = N->getOperand(0);
8225 switch (N->getOpcode()) {
8227 assert(0 && "Unknown shift opcode!");
8230 if (VT == MVT::v2i64)
8231 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8232 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8234 if (VT == MVT::v4i32)
8235 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8236 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8238 if (VT == MVT::v8i16)
8239 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8240 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8244 if (VT == MVT::v4i32)
8245 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8246 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8248 if (VT == MVT::v8i16)
8249 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8250 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8254 if (VT == MVT::v2i64)
8255 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8256 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8258 if (VT == MVT::v4i32)
8259 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8260 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8262 if (VT == MVT::v8i16)
8263 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8264 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8271 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8272 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8273 const X86Subtarget *Subtarget) {
8274 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8275 // the FP state in cases where an emms may be missing.
8276 // A preferable solution to the general problem is to figure out the right
8277 // places to insert EMMS. This qualifies as a quick hack.
8279 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8280 StoreSDNode *St = cast<StoreSDNode>(N);
8281 MVT VT = St->getValue().getValueType();
8282 if (VT.getSizeInBits() != 64)
8285 const Function *F = DAG.getMachineFunction().getFunction();
8286 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8287 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8288 && Subtarget->hasSSE2();
8289 if ((VT.isVector() ||
8290 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8291 isa<LoadSDNode>(St->getValue()) &&
8292 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8293 St->getChain().hasOneUse() && !St->isVolatile()) {
8294 SDNode* LdVal = St->getValue().getNode();
8296 int TokenFactorIndex = -1;
8297 SmallVector<SDValue, 8> Ops;
8298 SDNode* ChainVal = St->getChain().getNode();
8299 // Must be a store of a load. We currently handle two cases: the load
8300 // is a direct child, and it's under an intervening TokenFactor. It is
8301 // possible to dig deeper under nested TokenFactors.
8302 if (ChainVal == LdVal)
8303 Ld = cast<LoadSDNode>(St->getChain());
8304 else if (St->getValue().hasOneUse() &&
8305 ChainVal->getOpcode() == ISD::TokenFactor) {
8306 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8307 if (ChainVal->getOperand(i).getNode() == LdVal) {
8308 TokenFactorIndex = i;
8309 Ld = cast<LoadSDNode>(St->getValue());
8311 Ops.push_back(ChainVal->getOperand(i));
8315 if (!Ld || !ISD::isNormalLoad(Ld))
8318 // If this is not the MMX case, i.e. we are just turning i64 load/store
8319 // into f64 load/store, avoid the transformation if there are multiple
8320 // uses of the loaded value.
8321 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8324 DebugLoc LdDL = Ld->getDebugLoc();
8325 DebugLoc StDL = N->getDebugLoc();
8326 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8327 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8329 if (Subtarget->is64Bit() || F64IsLegal) {
8330 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8331 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8332 Ld->getBasePtr(), Ld->getSrcValue(),
8333 Ld->getSrcValueOffset(), Ld->isVolatile(),
8334 Ld->getAlignment());
8335 SDValue NewChain = NewLd.getValue(1);
8336 if (TokenFactorIndex != -1) {
8337 Ops.push_back(NewChain);
8338 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8341 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8342 St->getSrcValue(), St->getSrcValueOffset(),
8343 St->isVolatile(), St->getAlignment());
8346 // Otherwise, lower to two pairs of 32-bit loads / stores.
8347 SDValue LoAddr = Ld->getBasePtr();
8348 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8349 DAG.getConstant(4, MVT::i32));
8351 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8352 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8353 Ld->isVolatile(), Ld->getAlignment());
8354 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8355 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8357 MinAlign(Ld->getAlignment(), 4));
8359 SDValue NewChain = LoLd.getValue(1);
8360 if (TokenFactorIndex != -1) {
8361 Ops.push_back(LoLd);
8362 Ops.push_back(HiLd);
8363 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8367 LoAddr = St->getBasePtr();
8368 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8369 DAG.getConstant(4, MVT::i32));
8371 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8372 St->getSrcValue(), St->getSrcValueOffset(),
8373 St->isVolatile(), St->getAlignment());
8374 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8376 St->getSrcValueOffset() + 4,
8378 MinAlign(St->getAlignment(), 4));
8379 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8384 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8385 /// X86ISD::FXOR nodes.
8386 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8387 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8388 // F[X]OR(0.0, x) -> x
8389 // F[X]OR(x, 0.0) -> x
8390 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8391 if (C->getValueAPF().isPosZero())
8392 return N->getOperand(1);
8393 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8394 if (C->getValueAPF().isPosZero())
8395 return N->getOperand(0);
8399 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8400 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8401 // FAND(0.0, x) -> 0.0
8402 // FAND(x, 0.0) -> 0.0
8403 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8404 if (C->getValueAPF().isPosZero())
8405 return N->getOperand(0);
8406 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8407 if (C->getValueAPF().isPosZero())
8408 return N->getOperand(1);
8412 static SDValue PerformBTCombine(SDNode *N,
8414 TargetLowering::DAGCombinerInfo &DCI) {
8415 // BT ignores high bits in the bit index operand.
8416 SDValue Op1 = N->getOperand(1);
8417 if (Op1.hasOneUse()) {
8418 unsigned BitWidth = Op1.getValueSizeInBits();
8419 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8420 APInt KnownZero, KnownOne;
8421 TargetLowering::TargetLoweringOpt TLO(DAG);
8422 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8423 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8424 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8425 DCI.CommitTargetLoweringOpt(TLO);
8430 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
8431 SDValue Op = N->getOperand(0);
8432 if (Op.getOpcode() == ISD::BIT_CONVERT)
8433 Op = Op.getOperand(0);
8434 MVT VT = N->getValueType(0), OpVT = Op.getValueType();
8435 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
8436 VT.getVectorElementType().getSizeInBits() ==
8437 OpVT.getVectorElementType().getSizeInBits()) {
8438 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
8443 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8444 DAGCombinerInfo &DCI) const {
8445 SelectionDAG &DAG = DCI.DAG;
8446 switch (N->getOpcode()) {
8448 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8449 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8450 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8451 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8454 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8455 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8457 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8458 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8459 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8460 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
8466 //===----------------------------------------------------------------------===//
8467 // X86 Inline Assembly Support
8468 //===----------------------------------------------------------------------===//
8470 /// getConstraintType - Given a constraint letter, return the type of
8471 /// constraint it is for this target.
8472 X86TargetLowering::ConstraintType
8473 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8474 if (Constraint.size() == 1) {
8475 switch (Constraint[0]) {
8487 return C_RegisterClass;
8495 return TargetLowering::getConstraintType(Constraint);
8498 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8499 /// with another that has more specific requirements based on the type of the
8500 /// corresponding operand.
8501 const char *X86TargetLowering::
8502 LowerXConstraint(MVT ConstraintVT) const {
8503 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8504 // 'f' like normal targets.
8505 if (ConstraintVT.isFloatingPoint()) {
8506 if (Subtarget->hasSSE2())
8508 if (Subtarget->hasSSE1())
8512 return TargetLowering::LowerXConstraint(ConstraintVT);
8515 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8516 /// vector. If it is invalid, don't add anything to Ops.
8517 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8520 std::vector<SDValue>&Ops,
8521 SelectionDAG &DAG) const {
8522 SDValue Result(0, 0);
8524 switch (Constraint) {
8527 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8528 if (C->getZExtValue() <= 31) {
8529 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8535 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8536 if (C->getZExtValue() <= 63) {
8537 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8543 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8544 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
8545 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8551 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8552 if (C->getZExtValue() <= 255) {
8553 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8559 // 32-bit signed value
8560 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8561 const ConstantInt *CI = C->getConstantIntValue();
8562 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8563 // Widen to 64 bits here to get it sign extended.
8564 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8567 // FIXME gcc accepts some relocatable values here too, but only in certain
8568 // memory models; it's complicated.
8573 // 32-bit unsigned value
8574 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8575 const ConstantInt *CI = C->getConstantIntValue();
8576 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8577 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8581 // FIXME gcc accepts some relocatable values here too, but only in certain
8582 // memory models; it's complicated.
8586 // Literal immediates are always ok.
8587 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8588 // Widen to 64 bits here to get it sign extended.
8589 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8593 // If we are in non-pic codegen mode, we allow the address of a global (with
8594 // an optional displacement) to be used with 'i'.
8595 GlobalAddressSDNode *GA = 0;
8598 // Match either (GA), (GA+C), (GA+C1+C2), etc.
8600 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
8601 Offset += GA->getOffset();
8603 } else if (Op.getOpcode() == ISD::ADD) {
8604 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8605 Offset += C->getZExtValue();
8606 Op = Op.getOperand(0);
8609 } else if (Op.getOpcode() == ISD::SUB) {
8610 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8611 Offset += -C->getZExtValue();
8612 Op = Op.getOperand(0);
8617 // Otherwise, this isn't something we can handle, reject it.
8622 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(), Offset, DAG);
8624 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8631 if (Result.getNode()) {
8632 Ops.push_back(Result);
8635 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8639 std::vector<unsigned> X86TargetLowering::
8640 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8642 if (Constraint.size() == 1) {
8643 // FIXME: not handling fp-stack yet!
8644 switch (Constraint[0]) { // GCC X86 Constraint Letters
8645 default: break; // Unknown constraint letter
8646 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8649 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8650 else if (VT == MVT::i16)
8651 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8652 else if (VT == MVT::i8)
8653 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8654 else if (VT == MVT::i64)
8655 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8660 return std::vector<unsigned>();
8663 std::pair<unsigned, const TargetRegisterClass*>
8664 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8666 // First, see if this is a constraint that directly corresponds to an LLVM
8668 if (Constraint.size() == 1) {
8669 // GCC Constraint Letters
8670 switch (Constraint[0]) {
8672 case 'r': // GENERAL_REGS
8673 case 'R': // LEGACY_REGS
8674 case 'l': // INDEX_REGS
8676 return std::make_pair(0U, X86::GR8RegisterClass);
8678 return std::make_pair(0U, X86::GR16RegisterClass);
8679 if (VT == MVT::i32 || !Subtarget->is64Bit())
8680 return std::make_pair(0U, X86::GR32RegisterClass);
8681 return std::make_pair(0U, X86::GR64RegisterClass);
8682 case 'f': // FP Stack registers.
8683 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8684 // value to the correct fpstack register class.
8685 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
8686 return std::make_pair(0U, X86::RFP32RegisterClass);
8687 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
8688 return std::make_pair(0U, X86::RFP64RegisterClass);
8689 return std::make_pair(0U, X86::RFP80RegisterClass);
8690 case 'y': // MMX_REGS if MMX allowed.
8691 if (!Subtarget->hasMMX()) break;
8692 return std::make_pair(0U, X86::VR64RegisterClass);
8693 case 'Y': // SSE_REGS if SSE2 allowed
8694 if (!Subtarget->hasSSE2()) break;
8696 case 'x': // SSE_REGS if SSE1 allowed
8697 if (!Subtarget->hasSSE1()) break;
8699 switch (VT.getSimpleVT()) {
8701 // Scalar SSE types.
8704 return std::make_pair(0U, X86::FR32RegisterClass);
8707 return std::make_pair(0U, X86::FR64RegisterClass);
8715 return std::make_pair(0U, X86::VR128RegisterClass);
8721 // Use the default implementation in TargetLowering to convert the register
8722 // constraint into a member of a register class.
8723 std::pair<unsigned, const TargetRegisterClass*> Res;
8724 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8726 // Not found as a standard register?
8727 if (Res.second == 0) {
8728 // GCC calls "st(0)" just plain "st".
8729 if (StringsEqualNoCase("{st}", Constraint)) {
8730 Res.first = X86::ST0;
8731 Res.second = X86::RFP80RegisterClass;
8733 // 'A' means EAX + EDX.
8734 if (Constraint == "A") {
8735 Res.first = X86::EAX;
8736 Res.second = X86::GRADRegisterClass;
8741 // Otherwise, check to see if this is a register class of the wrong value
8742 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
8743 // turn into {ax},{dx}.
8744 if (Res.second->hasType(VT))
8745 return Res; // Correct type already, nothing to do.
8747 // All of the single-register GCC register classes map their values onto
8748 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
8749 // really want an 8-bit or 32-bit register, map to the appropriate register
8750 // class and return the appropriate register.
8751 if (Res.second == X86::GR16RegisterClass) {
8752 if (VT == MVT::i8) {
8753 unsigned DestReg = 0;
8754 switch (Res.first) {
8756 case X86::AX: DestReg = X86::AL; break;
8757 case X86::DX: DestReg = X86::DL; break;
8758 case X86::CX: DestReg = X86::CL; break;
8759 case X86::BX: DestReg = X86::BL; break;
8762 Res.first = DestReg;
8763 Res.second = X86::GR8RegisterClass;
8765 } else if (VT == MVT::i32) {
8766 unsigned DestReg = 0;
8767 switch (Res.first) {
8769 case X86::AX: DestReg = X86::EAX; break;
8770 case X86::DX: DestReg = X86::EDX; break;
8771 case X86::CX: DestReg = X86::ECX; break;
8772 case X86::BX: DestReg = X86::EBX; break;
8773 case X86::SI: DestReg = X86::ESI; break;
8774 case X86::DI: DestReg = X86::EDI; break;
8775 case X86::BP: DestReg = X86::EBP; break;
8776 case X86::SP: DestReg = X86::ESP; break;
8779 Res.first = DestReg;
8780 Res.second = X86::GR32RegisterClass;
8782 } else if (VT == MVT::i64) {
8783 unsigned DestReg = 0;
8784 switch (Res.first) {
8786 case X86::AX: DestReg = X86::RAX; break;
8787 case X86::DX: DestReg = X86::RDX; break;
8788 case X86::CX: DestReg = X86::RCX; break;
8789 case X86::BX: DestReg = X86::RBX; break;
8790 case X86::SI: DestReg = X86::RSI; break;
8791 case X86::DI: DestReg = X86::RDI; break;
8792 case X86::BP: DestReg = X86::RBP; break;
8793 case X86::SP: DestReg = X86::RSP; break;
8796 Res.first = DestReg;
8797 Res.second = X86::GR64RegisterClass;
8800 } else if (Res.second == X86::FR32RegisterClass ||
8801 Res.second == X86::FR64RegisterClass ||
8802 Res.second == X86::VR128RegisterClass) {
8803 // Handle references to XMM physical registers that got mapped into the
8804 // wrong class. This can happen with constraints like {xmm0} where the
8805 // target independent register mapper will just pick the first match it can
8806 // find, ignoring the required type.
8808 Res.second = X86::FR32RegisterClass;
8809 else if (VT == MVT::f64)
8810 Res.second = X86::FR64RegisterClass;
8811 else if (X86::VR128RegisterClass->hasType(VT))
8812 Res.second = X86::VR128RegisterClass;
8818 //===----------------------------------------------------------------------===//
8819 // X86 Widen vector type
8820 //===----------------------------------------------------------------------===//
8822 /// getWidenVectorType: given a vector type, returns the type to widen
8823 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
8824 /// If there is no vector type that we want to widen to, returns MVT::Other
8825 /// When and where to widen is target dependent based on the cost of
8826 /// scalarizing vs using the wider vector type.
8828 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
8829 assert(VT.isVector());
8830 if (isTypeLegal(VT))
8833 // TODO: In computeRegisterProperty, we can compute the list of legal vector
8834 // type based on element type. This would speed up our search (though
8835 // it may not be worth it since the size of the list is relatively
8837 MVT EltVT = VT.getVectorElementType();
8838 unsigned NElts = VT.getVectorNumElements();
8840 // On X86, it make sense to widen any vector wider than 1
8844 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
8845 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
8846 MVT SVT = (MVT::SimpleValueType)nVT;
8848 if (isTypeLegal(SVT) &&
8849 SVT.getVectorElementType() == EltVT &&
8850 SVT.getVectorNumElements() > NElts)